A three-dimensional model design optimization method and system based on deep learning

Abstract

The application relates to the technical field of architectural design, in particular to a three-dimensional model design optimization method and system based on deep learning. The method comprises the following steps: obtaining original three-dimensional model data; performing data preprocessing on the obtained three-dimensional model data; constructing an intelligent analysis engine based on the preprocessed three-dimensional model data; performing parameter optimization according to the constructed intelligent analysis engine, wherein the output result of the intelligent analysis engine is subjected to multi-objective optimization through an optimization algorithm; performing prediction simulation by using the optimized intelligent analysis engine; setting simulation parameters to perform performance prediction; and performing real-time sharing and intelligent analysis integration of the prediction simulation result. The deep learning module of the intelligent analysis engine can adapt to different types of three-dimensional model data, can perform performance simulation and evaluation on three-dimensional designs in different fields, and improves the universality and adaptability of the design method.

Description

Technical Field

[0001] This invention relates to the field of architectural design technology, and in particular to a method and system for optimizing 3D model design based on deep learning. Background Technology

[0002] In the field of architectural design, with the development of the times and the increasing scale and complexity of projects, traditional 3D architectural design methods have gradually exposed many drawbacks. Traditional 3D architectural design mainly relies on the designer's personal experience and manual operation. Designers often need to spend a lot of time and energy building 3D architectural models. In this process, they evaluate the feasibility and performance of the design through manual calculation and analysis. For example, for the stability analysis of building structures, designers may need to perform complex calculations based on empirical formulas, which is not only inefficient but also prone to errors due to human factors.

[0003] Traditional methods have limitations in showcasing a building's exterior and internal structure. They struggle to present the building's details comprehensively and accurately, and fail to allow relevant personnel to identify problems in the design, such as unreasonable spatial layout or insufficient structural stability, in advance. For example, traditional 3D model displays may not clearly show the spatial relationships between different floors, leading to the discovery of unreasonable staircase placements during actual construction. This results in frequent changes and adjustments, significantly increasing project costs and timelines.

[0004] Furthermore, traditional design methods are not precise enough in terms of building energy consumption analysis and environmental impact assessment. Assessments of building performance such as lighting, ventilation, and insulation often rely on empirical judgment or simple calculation tools, which are insufficient to meet the requirements of sustainable development in modern buildings. For example, when assessing building energy consumption, it may not accurately account for the impact of changes in sunlight and temperature at different seasons and times on energy demand. Currently, there is a need for a deep learning-based three-dimensional model design optimization method and system. Summary of the Invention

[0005] To address the significant shortcomings of traditional 3D design methods in terms of efficiency, accuracy, collaboration, and adaptability, this invention provides a 3D model design optimization method and system based on deep learning.

[0006] Firstly, the present invention provides a deep learning-based method for designing and optimizing 3D models, employing the following technical solution: A deep learning-based method for designing and optimizing 3D models, comprising: Obtain the original 3D model data; The acquired 3D model data is preprocessed, including data cleaning, and the cleaned 3D model data is then converted in format and its features are extracted. The intelligent analysis engine is built based on the preprocessed 3D model data, including the construction of deep learning modules, machine learning modules and data mining modules; The parameters of the completed intelligent analysis engine are optimized, including multi-objective optimization of the output of the intelligent analysis engine through optimization algorithms; The optimized intelligent analysis engine is used for predictive simulation, including integrated structural analysis software, energy consumption simulation software and lighting simulation software, to simulate model performance and set simulation parameters for performance prediction. Integrate and share the prediction simulation results in real time with intelligent analysis.

[0007] Furthermore, the data cleaning process involves format conversion and feature extraction of the cleaned 3D model data. This includes using a Gaussian filtering algorithm to denoise the 3D model data, then using linear interpolation to complete missing or incomplete data, converting 3D model data from different sources into a unified internal format, and finally using computer graphics algorithms to extract the geometric features of the model, combined with a material library and physical property calculation tools to extract the physical features of the model.

[0008] Furthermore, the construction of the deep learning module, machine learning module, and data mining module includes constructing a hybrid model of convolutional neural networks and recurrent neural networks. The convolutional neural network is used to extract the spatial features of the 3D model, the recurrent neural network is used to process the temporal information of the model, the analysis results are collected to construct a training dataset, and transfer learning and reinforcement learning techniques are used to improve the generalization ability of the model.

[0009] Furthermore, the construction of the deep learning module, machine learning module, and data mining module also includes using support vector machines and random forests to learn classification algorithms to classify and evaluate the performance of the model, and using regression algorithms to predict the performance indicators of the model after classification.

[0010] Furthermore, the construction of the deep learning module, machine learning module, and data mining module also includes calculating the association rules between model parameters using the Apriori algorithm, and finally performing K-means clustering analysis based on the design case to classify potential design patterns and optimization strategies.

[0011] Furthermore, the multi-objective optimization of the output of the intelligent analysis engine through the optimization algorithm includes determining the direction and objective of optimization based on the output of the intelligent analysis engine, iterating the optimization algorithm using a genetic algorithm, balancing multiple objectives by setting appropriate objective functions and constraints, calculating the fitness value of each solution based on the objective function, and determining the final number of iterations based on the fitness value.

[0012] Furthermore, the prediction simulation using the optimized intelligent analysis engine includes importing the optimized 3D model data into structural analysis software, determining the range and values ​​of simulation parameters according to the optimization objective, performing multi-parameter combination simulation based on the values ​​of the simulation parameters, and using analysis software to perform prediction analysis based on the set simulation parameters.

[0013] Secondly, a deep learning-based 3D model design optimization system includes: The data acquisition module is configured to acquire raw 3D model data. The preprocessing module is configured to: perform data preprocessing on the acquired 3D model data, including data cleaning, and perform format conversion and feature extraction on the cleaned 3D model data; The conversion module is configured to build an intelligent analysis engine based on the preprocessed 3D model data, including building a deep learning module, a machine learning module, and a data mining module. The optimization module is configured to: optimize parameters based on the constructed intelligent analysis engine, including performing multi-objective optimization on the output of the intelligent analysis engine through optimization algorithms; The prediction module is configured to: perform predictive simulations using an optimized intelligent analysis engine, including integrated structural analysis software, energy consumption simulation software, and illumination simulation software; perform model performance simulations; and set simulation parameters to predict performance. The transformation module is configured to integrate and intelligently analyze the prediction simulation results in real time.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned deep learning-based 3D model design optimization method.

[0015] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the deep learning-based 3D model design optimization method.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This invention utilizes the deep learning module of an intelligent analysis engine to adapt to different types of 3D model data. Through transfer learning and reinforcement learning techniques, it continuously learns and adapts to new design tasks and scenarios. During the multi-objective optimization process, it flexibly adjusts the objective function and constraints according to different design requirements and optimization objectives, enabling the design method to adapt to various complex design requirements. Furthermore, it integrates multiple professional software programs, allowing for performance simulation and evaluation of 3D designs in different fields, thereby improving the versatility and adaptability of the design method.

[0017] 2. This invention achieves multi-faceted performance simulation by importing optimized 3D model data into structural analysis software, energy consumption simulation software, and lighting simulation software, ensuring the reliability and rationality of the design, reducing human error, determining the range and values ​​of simulation parameters according to the optimization target, and performing multi-parameter combination simulations. This allows for the exploration of the performance of different design schemes, helping designers find the optimal combination of design parameters and improve design quality. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of a deep learning-based 3D model design and optimization method according to an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings.

[0020] Example 1 Reference Figure 1 This embodiment of a deep learning-based 3D model design optimization method includes: Obtain the original 3D model data; The acquired 3D model data is preprocessed, including data cleaning, and the cleaned 3D model data is then converted in format and its features are extracted. The intelligent analysis engine is built based on the preprocessed 3D model data, including the construction of deep learning modules, machine learning modules and data mining modules; The parameters of the completed intelligent analysis engine are optimized, including multi-objective optimization of the output of the intelligent analysis engine through optimization algorithms; The optimized intelligent analysis engine is used for predictive simulation, including integrated structural analysis software, energy consumption simulation software and lighting simulation software, to simulate model performance and set simulation parameters for performance prediction. Integrate and share the prediction simulation results in real time with intelligent analysis.

[0021] Specifically, a deep learning-based method for designing and optimizing 3D models includes the following steps: like Figure 1As shown, S1, obtain the original 3D model data; First, a large amount of raw data related to a specific field is collected as input. This data includes various parameters of the 3D design model, performance indicators, historical design cases, etc. In the field of architectural design, the data includes the building's dimensions, material information, energy consumption data, lighting conditions, etc.

[0022] S2. Perform data preprocessing on the acquired 3D model data, including data cleaning, and perform format conversion and feature extraction on the cleaned 3D model data; First, the acquired data is denoised using a Gaussian filtering algorithm. The basic principle is to determine the filter weights based on the shape of the Gaussian function. For a two-dimensional image (the three-dimensional model data can be viewed as a combination of a series of two-dimensional slices), the pixel value at coordinates is... The value after Gaussian filtering can be calculated using the following formula: , in, Represented in pixels, Expressed as standard deviation, it determines the smoothness of the filter. This indicates that after Gaussian filtering, the coordinates are... Pixel value at that location, This indicates that the original image is located at coordinates... Pixel value at that location, This represents the normalization coefficient, which is then used to complete missing or incomplete data using linear interpolation. Assuming two data points are known... and To obtain the value by interpolation at a certain point, the linear interpolation formula is: , in, This represents the x-coordinate of the position to be interpolated. and Represented as the x-coordinates of two known data points. and They are respectively represented as corresponding to and The ordinate of the data points This is represented as the value obtained by interpolation at the x-coordinate. Finally, a format conversion tool is used to convert 3D model data from different sources (such as RVT, STEP, IGES, etc.) into a unified internal format. Computer graphics algorithms are used to extract the geometric features of the model, such as volume, surface area, and shape parameters (aspect ratio, roundness, etc.). Volume can be calculated by integration, surface area can be calculated by summing the areas of the triangles on the model surface, and shape parameters are calculated using corresponding formulas. For example, for a 3D model, its maximum and minimum dimensions in the three coordinate axes can be determined first. Assume that the maximum dimension in the x-axis direction is maxX and the minimum dimension is minX; the maximum dimension in the y-axis direction is maxY and the minimum dimension is minY; and the maximum dimension in the z-axis direction is maxZ and the minimum dimension is minZ. The aspect ratio can be defined as max(length) / min(length), where length can be the ratio of the maximum to the minimum value among maxX - minX, maxY - minY, or maxZ - minZ. Combining the material library and physical property calculation tools, the physical features of the model can be extracted, such as material density, thermal conductivity, and structural strength. Material density can be obtained from the material library, while thermal conductivity and structural strength can be calculated using the corresponding physical calculation formulas.

[0023] S3. Construct an intelligent analysis engine based on the preprocessed 3D model data, including building a deep learning module, a machine learning module, and a data mining module; The acquired raw data is fed into the deep learning module to construct a hybrid model of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). CNN is used to extract spatial features of the 3D model, such as the shape and structure of objects; RNN is used to process the model's temporal information, such as trends during the design process. Backpropagation is used to adjust model parameters, and the model is trained using a training dataset constructed from numerous design cases and analysis results. Simultaneously, transfer learning and reinforcement learning techniques are employed to improve the model's generalization ability and adaptability. Specifically, CNN automatically learns the spatial features of images (2D slices of the 3D model) through multiple convolutional and pooling layers. The operation of a convolutional layer can be viewed as performing a convolution operation between the input image and a set of learnable convolutional kernels. For a 2D image, the output value O(x,y) after convolution at coordinates (x,y) can be calculated using the following formula: , Where I(x,y) is the pixel value of the input image, and K(i,j) is the value of the convolution kernel. Pooling layers typically use max pooling or average pooling to reduce the resolution of the feature map while preserving the main features. For example, max pooling takes the maximum value within a small region as the output. Then, a recurrent neural network (RNN) is constructed to process temporal information. RNNs are suitable for processing sequential data and can be used in 3D models to handle information that changes over time. The basic unit of an RNN is a neuron, and its output depends not only on the current input but also on the hidden state from the previous time step, as shown in the formula: , in, It is the input at the current moment. It is the hidden state at the current moment. and It is a weight matrix. This refers to the bias term. Transfer learning leverages knowledge learned from other related tasks to accelerate learning in the current task. For example, the weights of a CNN model pre-trained on a large-scale image dataset can be used as initial values ​​and applied to a 3D model feature extraction task. Then, the feature vectors output from the deep learning stage are used as input to the machine learning module. Specifically, classification algorithms such as Support Vector Machine (SVM) and Random Forest (RF) are applied to classify and evaluate model performance, determining its merits and applicable scenarios. Simultaneously, regression algorithms such as linear regression and ridge regression are used to predict model performance metrics, such as energy consumption and illumination. SVM classifies data by finding an optimal hyperplane. For binary classification problems, assuming the training data is... ,in, It is the input feature vector. These are the class labels. The goal of SVM is to find a hyperplane that maximizes the distance between the two classes of data points and the hyperplane. RF is an ensemble learning algorithm composed of multiple decision trees. For an input sample, each decision tree provides a classification result, and the final classification result is determined by majority vote. Decision trees are constructed by splitting features. At each node, an optimal feature and split point are selected to maximize the purity of the resulting child nodes. Purity can be measured using metrics such as information entropy and the Gini index. Linear regression is used to predict continuous numerical target variables, estimating regression coefficients by minimizing the mean squared error (MSE). Ridge regression adds a regularization term to linear regression to prevent overfitting. Using raw 3D model data, intermediate results from deep learning and machine learning as input, the Apriori algorithm or FP-Growth algorithm is used to mine the association rules between design parameters, discovering potential design patterns and relationships. K-means clustering or hierarchical clustering analysis is performed on similar design cases to better understand the design space and discover optimization strategies.

[0024] Finally, the data mining module uses the Apriori algorithm to mine frequent itemsets and association rules. A frequent itemset is a set of items that appear frequently in a dataset. The algorithm iterates continuously, first finding frequent 1-itemsets, then generating frequent 2-itemsets based on them, and so on, until no larger frequent itemsets can be found. The association rule takes the form of... In this context, A and B are itemsets, and support and confidence are two metrics that measure the strength of association rules. Support refers to the proportion of transactions that contain both A and B in the total number of transactions; confidence refers to the proportion of transactions that contain both A and B that also contain B.

[0025] Then, K initial cluster centers are randomly selected, and each data point is assigned to the cluster containing the nearest cluster center. Next, the center of each cluster is recalculated, and this process is repeated until the cluster centers no longer change or the maximum number of iterations is reached.

[0026] S4. Optimize the parameters of the completed intelligent analysis engine, including multi-objective optimization of the output results of the intelligent analysis engine through optimization algorithms; Based on the output of the intelligent analysis engine, the direction and objectives of optimization are determined. For example, for material selection optimization, appropriate material combinations can be selected according to the model's performance requirements and cost constraints; for structural layout optimization, the structural characteristics of the model can be analyzed to find a more reasonable layout scheme to improve strength and stability; for lighting design optimization, light source parameters and model surface materials can be adjusted to improve lighting effects. A genetic algorithm (GA) is used to find the optimal solution in the search space by simulating processes such as natural selection, crossover, and mutation. Conflicts between different design objectives, such as cost, energy consumption, and aesthetics, are considered. Through the set objective function and constraints, the optimization algorithm weighs multiple objectives. The objective function is expressed as: , Where C represents the target cost, E represents energy consumption, and B represents aesthetics. The design parameters are represented as weight coefficients, and a design scheme with optimal overall performance is found. The design parameters are encoded; for genetic algorithms, these can be encoded as chromosomes. Through selection, crossover, and mutation, the optimization process iterates continuously, gradually approaching the optimal solution. In each iteration, a superior design scheme is selected based on the fitness value, and genetic operations are performed to generate a new design scheme. A maximum number of iterations T is pre-set. When the current iteration number t reaches T, the optimization process stops. When the stopping condition is met, the optimization process stops, and the optimal combination of design parameters and performance indicators are output.

[0027] S5. Utilize the optimized intelligent analysis engine to perform predictive simulations, including integrated structural analysis software, energy consumption simulation software, and illumination simulation software; perform model performance simulations and set simulation parameters to predict performance. 1. Integrate structural analysis software: such as ANSYS, import the optimized 3D model data into the structural analysis software to evaluate the model's structural strength, stability, and deformation. By setting different load conditions and boundary constraints, the stress conditions of the model in actual use can be simulated, potential problems can be predicted, and a basis for structural optimization can be provided.

[0028] 2. Integrate energy consumption simulation software: such as EnergyPlus, to analyze the energy consumption performance of the model. Based on factors such as the model's geometry, material properties, and lighting conditions, energy consumption simulation software can calculate the energy consumption of the model under different operating conditions, such as electricity and heat. By optimizing design parameters, the model's energy consumption can be reduced, and energy utilization efficiency can be improved.

[0029] 3. Integrate lighting simulation software, such as Radiance, to simulate the lighting effects on the model. Different light source positions, intensities, and colors, as well as the model's surface materials and reflective properties, can be set to evaluate indicators such as lighting uniformity, brightness, and color reproduction. The lighting simulation results can provide a reference for lighting design, improving the model's visual appeal and comfort.

[0030] The specific integration method involves developing a data interface to ensure data exchange and interaction between different simulation software. This data interface imports the geometric data of the 3D model from the design software into the structural analysis software. After structural analysis, the results are then passed to the energy consumption simulation software and the lighting simulation software. For the analysis model, performance indicators can be obtained by calling its API or reading the output file for comprehensive performance simulation. Since different simulation software may support different model formats, model conversion is required to ensure that the 3D model can be correctly loaded and run in each simulation software.

[0031] Based on design requirements and optimization objectives, the range and values ​​of simulation parameters are determined. For example, in structural analysis, different load magnitudes, directions, and points of application can be set; in energy consumption simulation, factors such as indoor temperature, humidity, and human activity can be adjusted; in lighting simulation, parameters such as the position, intensity, and color of the light source can be changed. Considering the interactions between different parameters, multi-parameter combination simulations are performed. By adjusting the values ​​of multiple parameters, the performance of different design schemes can be explored, and the optimal parameter combination can be found.

[0032] The integrated professional software performs performance predictions. Based on the set simulation parameters, the software calculates and analyzes the model, generating corresponding performance index results, such as structural stress, deformation, energy consumption, and light intensity. The performance prediction results are analyzed to evaluate the merits of the design scheme. Different design schemes can be compared in terms of various performance indicators to determine the optimal design scheme. Simultaneously, the analysis results can also reveal problems in the design and potential improvement directions, providing further optimization suggestions based on the performance prediction results.

[0033] S6. Real-time sharing and intelligent analysis integration of prediction and simulation results.

[0034] Integrate version control mechanisms, such as Git, to manage design changes and history. Git supports branch management and merging operations by recording version information for each commit. Integrate the intelligent analysis engine into the collaborative work platform. Real-time intelligent analysis services are achieved by calling the intelligent analysis engine's interface.

[0035] Example 2 The difference between this embodiment and Embodiment 1 is that this embodiment provides a deep learning-based 3D model design optimization system, including: The data acquisition module is configured to acquire raw 3D model data. The preprocessing module is configured to: perform data preprocessing on the acquired 3D model data, including data cleaning, and perform format conversion and feature extraction on the cleaned 3D model data; The conversion module is configured to build an intelligent analysis engine based on the preprocessed 3D model data, including building a deep learning module, a machine learning module, and a data mining module. The optimization module is configured to: optimize parameters based on the constructed intelligent analysis engine, including performing multi-objective optimization on the output of the intelligent analysis engine through optimization algorithms; The prediction module is configured to: perform predictive simulations using an optimized intelligent analysis engine, including integrated structural analysis software, energy consumption simulation software, and illumination simulation software; perform model performance simulations; and set simulation parameters to predict performance. The transformation module is configured to integrate and intelligently analyze the prediction simulation results in real time.

[0036] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the described deep learning-based 3D model design optimization method.

[0037] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions adapted for loading and execution by the processor of the described deep learning-based 3D model design optimization method.

[0038] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for designing and optimizing 3D models based on deep learning, characterized in that, include: Obtain the original 3D model data; The three-dimensional model data is preprocessed by treating it as a combination of a series of two-dimensional slices. Gaussian filtering is used to denoise the two-dimensional slices, and linear interpolation is used to complete missing or incomplete data. Three-dimensional model data from different sources in RVT, STEP, and IGES formats are converted into a unified internal format. Geometric features, including volume, surface area, and shape parameters, are extracted using computer graphics algorithms, and physical features, including material density, thermal conductivity, and structural strength, are extracted using a material library and physical property calculation tools. An intelligent analysis engine is built based on the preprocessed 3D model data. The intelligent analysis engine includes a deep learning module, a machine learning module, and a data mining module. The deep learning module constructs a hybrid model of convolutional neural networks and recurrent neural networks. The convolutional neural network extracts spatial features from two-dimensional slices of the three-dimensional model, while the recurrent neural network processes temporal information in the three-dimensional model design process. The machine learning module uses support vector machines and random forests to classify and evaluate model performance, and uses regression algorithms to predict model performance metrics. The data mining module takes the original 3D model data, intermediate results of deep learning and machine learning as input, mines the association rules between design parameters through the Apriori algorithm, and performs K-means clustering analysis on similar design cases. The optimization direction and target of the architectural 3D model design parameters are determined based on the output of the intelligent analysis engine. The architectural 3D model design parameters include materials, structural layout, light source parameters, and model surface materials. The design parameters of the 3D building model are encoded into chromosomes of a genetic algorithm, and iterative processes are performed through selection, crossover, and mutation, according to the objective function. Calculate the fitness value of the proposed solution, where C represents the target cost, E represents energy consumption, and B represents aesthetics. Represented as weight coefficients, the optimization stops when the current iteration count reaches the preset maximum iteration count or the stopping condition is met, and the optimized design parameter combination and performance indicators are output. Predictive simulations are performed on the optimized 3D model. A data interface for data exchange and interaction between different simulation software is developed. The geometric data of the 3D model is imported from the design software into the structural analysis software for structural analysis, and the structural analysis results are transferred to the energy consumption simulation software and the lighting simulation software. Performance indicators including structural stress, deformation, energy consumption value and lighting intensity are obtained by calling the API of the corresponding simulation software or reading the output file. The range and value of simulation parameters are determined according to the optimization objectives, and multi-parameter combination simulations are performed. Integrate and share the prediction simulation results in real time with intelligent analysis.

2. A deep learning-based 3D model design optimization system, as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire raw 3D model data. The preprocessing module is configured to: perform data preprocessing on the acquired 3D model data, including data cleaning, and perform format conversion and feature extraction on the cleaned 3D model data; The conversion module is configured to build an intelligent analysis engine based on the preprocessed 3D model data, including building a deep learning module, a machine learning module, and a data mining module. The optimization module is configured to: optimize parameters based on the constructed intelligent analysis engine, including performing multi-objective optimization on the output of the intelligent analysis engine through optimization algorithms; The prediction module is configured to: perform predictive simulations using an optimized intelligent analysis engine, including integrated structural analysis software, energy consumption simulation software, and illumination simulation software; perform model performance simulations; and set simulation parameters to predict performance. The transformation module is configured to integrate and intelligently analyze the prediction simulation results in real time.

3. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded and executed by the processor of the terminal device as described in claim 1, which is a deep learning-based 3D model design optimization method.

4. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is configured to implement instructions, and the computer-readable storage medium is configured to store multiple instructions, characterized in that, The instructions are adapted to be loaded and executed by a processor as described in claim 1, which is a deep learning-based 3D model design optimization method.

Images