Three-dimensional modeling method and system based on exploration space learning and simulation optimization

Abstract

The application provides a three-dimensional modeling method and system based on exploration space learning and simulation optimization, and belongs to the technical field of three-dimensional design. The method specifically comprises the following steps: responding to user input demand information, matching or generating an initial three-dimensional model; based on the model representation type of the initial three-dimensional model, generating an exploration space of corresponding model parameters through neural network prediction; taking the highest score result output by a simulation simulator according to a preset scoring index as an optimization target, and combining a preset optimization strategy to iteratively search the model parameters in the exploration space to obtain an optimal model parameter scheme; and adjusting the initial three-dimensional model based on the optimal model parameter scheme to output a final three-dimensional model. The application automatically predicts the exploration space of model parameters through a neural network, and automatically iteratively optimizes the model parameters in combination with a simulation simulator, so that the artificial intervention in the modeling optimization process can be effectively reduced, the modeling efficiency is improved while the accuracy in a complex modeling scenario is ensured.

Description

Technical Field

[0001] This invention relates to the field of 3D design technology, and in particular to a 3D modeling method and system based on exploratory space learning and simulation optimization. Background Technology

[0002] 3D modeling technology is now widely used in many fields such as mechanical equipment, industrial product design, architectural structure, aerospace, and cultural and creative design. It is an important part of the entire process of product development, structural simulation, design iteration, and process verification. Currently, the mainstream 3D models are mainly divided into two types: parametric CAD models and patch mesh models. Parametric CAD models rely on structured parameters such as dimensions and angles to drive morphological changes and are mostly used for the design of regular models such as mechanical structures and standard parts. Patch mesh models, on the other hand, use vertices, patches, and control points to form geometric shapes, which can realize the design of free-form surfaces and complex shapes. Both types of models support the 3D design needs in different scenarios.

[0003] Traditional 3D modeling and design processes heavily rely on manual operation by designers. After building an initial 3D model based on design requirements, designers need to manually adjust various parameters of the parametric model or drag and drop key control points of the mesh model to modify its shape. Then, they use tools such as finite element analysis and mechanical simulation to verify the model's performance. The final design solution is obtained through repeated iterations of manual adjustment and simulation verification. This manual approach suffers from low design efficiency, and it is often difficult to balance multiple design objectives during manual design, resulting in modeling results that are difficult to adapt to design requirements.

[0004] While there are currently 3D design solutions that utilize parametric modeling tools and shape optimization methods to assist in model optimization and iteration in order to improve design efficiency, these methods still require users to manually set the initial model and the corresponding range of optimization variables and constraints. Although they can adapt to simple modeling needs, they are prone to getting stuck in local optima when faced with complex modeling needs, and the problems of low modeling efficiency and accuracy still exist. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of current 3D design methods, which require manual setting of the optimization basis for modeling, making it difficult to adapt to complex modeling needs and resulting in low modeling efficiency and accuracy. This invention provides a 3D modeling method and system based on exploration space learning and simulation optimization. It automatically matches or generates an initial 3D model by using requirement information, and combines this with a neural network to automatically predict the exploration space of model parameters, ensuring the accuracy of the initial parameter settings and avoiding getting stuck in local optima due to inaccurate manual parameter settings. Furthermore, it uses a simulation simulator to automatically iterate and optimize the model parameters, avoiding repetitive manual operations during the modeling optimization process. This ensures accuracy in complex modeling scenarios while improving modeling efficiency.

[0006] The objective of this invention is achieved through the following technical solution: 3D modeling methods based on exploratory spatial learning and simulation optimization include: Responding to user input, matching or generating an initial 3D model; Based on the model representation type of the initial 3D model, the exploration space corresponding to the model parameters is generated by predicting through a neural network; The optimization objective is to achieve the highest score output by the simulation simulator based on the preset scoring index. The optimal model parameter scheme is obtained by iteratively searching the model parameters in the exploration space in combination with the preset optimization strategy. The initial 3D model is adjusted based on the optimal model parameter scheme to output the final 3D model.

[0007] Furthermore, the initial 3D model is represented by either a parametric model representation or a patch mesh model representation.

[0008] Furthermore, when the model representation type of the initial 3D model is a parametric model representation, the exploration space for generating the corresponding model parameters is predicted by a neural network, including: Based on historical modeling data, the parameter sequence distribution and constraint relationship of model parameters are learned through neural networks, and the range of value changes of each model parameter is predicted, forming the exploration space of the corresponding model parameters.

[0009] Furthermore, the method of predicting the range of values ​​for each model parameter by learning the parameter sequence distribution and constraint relationships of the model parameters through a neural network based on historical modeling data includes: Numerical parameters of historical parameterized models are extracted from historical modeling data, corresponding parameter sequences are constructed, and parameter sequences of all historical parameterized models are integrated to obtain a training sample set. With the goal of fitting the distribution of parameter sequences in the training sample set, the neural network is trained to learn the parameter sequence distribution and constraint relationship of the model parameters based on the training sample set. The parameter sequence of the initial 3D model is input into the trained neural network. The neural network delineates the basic boundaries of each model parameter value based on the learned parameter sequence distribution, and corrects the basic boundaries of the values ​​in combination with the constraint relationship, thereby obtaining the range of value variation of each model parameter.

[0010] Furthermore, when the initial 3D model's model representation type is a patch mesh model representation, an exploration space for predicting and generating corresponding model parameters is generated through a neural network, including: Based on the initial 3D model, key control points are extracted, and combined with historical modeling data, the deformation displacement range of each key control point is predicted through a neural network, forming an exploration space for the corresponding model parameters.

[0011] Furthermore, the step of extracting key control points based on the initial 3D model and combining them with historical modeling data to predict the deformation displacement range of each key control point through a neural network, forming an exploration space for the corresponding model parameters, includes: Extract grid control points and corresponding historical deformation data from historical modeling data to construct a deformation training sample set; With the goal of fitting the deformation distribution of grid control points in the deformation training sample set, a neural network is trained based on the training sample set to learn the deformation distribution law and deformation constraint relationship of the grid control points. Key control points are extracted from the initial 3D model and input into the trained neural network. The neural network delineates the basic boundary of deformation for each key control point based on the learned deformation distribution law, and corrects the basic boundary of deformation in combination with deformation constraint relationship to obtain the deformation displacement range of each key control point.

[0012] Furthermore, the preset optimization strategy includes: Based on the initial 3D model, determine whether the current modeling satisfies the differentiability condition; If satisfied, a differentiable physical simulation simulator is constructed, and a gradient descent optimization strategy is matched. The optimization objective is to maximize the score output by the physical simulation simulator based on the preset scoring index, and the model parameters are iteratively searched in the exploration space. If the conditions are not met, a basic simulation simulator is constructed and a reinforcement learning optimization strategy is matched. The optimization objective is to maximize the score output by the basic simulation simulator based on the preset scoring index, and the model parameters are iteratively searched in the exploration space.

[0013] Furthermore, the construction of a differentiable physical simulation simulator, and the matching of a gradient descent optimization strategy, with the goal of maximizing the score output by the physical simulation simulator according to a preset scoring index, involves iteratively searching the model parameters within the exploration space, including: Obtain or generate a mesh representation of the initial 3D model, define geometric constraints and energy potential functions based on the mesh representation, and construct a differentiable physical simulation simulator; A comprehensive scoring function is constructed based on preset scoring indicators and their weights, and a loss function for iterative search is constructed based on the comprehensive scoring function. Input the current model parameters of the initial 3D model into the physical simulation simulator, and perform forward simulation to obtain the model state; Based on the model state obtained from the simulation, the comprehensive score and the corresponding loss function value are calculated, and the gradient of the loss function with respect to the current model parameters is calculated based on the differential calculation and the loss function value. Based on the calculated gradient, the model parameters are updated using a gradient descent optimization strategy. Among the current model parameters and the updated model parameters, save the model parameters with higher overall scores and determine whether the iteration termination condition is met; If the iteration termination condition is met, the optimal model parameter scheme is output based on the saved model parameters; If the iteration termination condition is not met, the saved model parameters are re-input into the physical simulation simulator as the current model parameters to optimize and update the model parameters until the iteration termination condition is met.

[0014] Furthermore, the construction of a basic simulation simulator and the matching of a reinforcement learning optimization strategy, with the goal of maximizing the score output by the basic simulation simulator according to a preset scoring metric, involves iteratively searching the model parameters within the exploration space, including: Construct a basic simulation simulator and build a comprehensive scoring function based on preset scoring indicators and their weights; The model parameters in the exploration space are encoded as state vectors, and the adjustment of model parameters is defined as actions. The comprehensive score is used as the reward, and a predefined reinforcement learning optimization strategy is used. The current model parameters of the initial 3D model are input into the basic simulation simulator for physical simulation, and the corresponding comprehensive score is calculated by combining the comprehensive scoring function. The action selection is optimized through reinforcement learning. Adjust the current model parameters based on the selected action, and obtain the updated model parameters; Re-enter the updated model parameters into the basic simulation simulator and obtain the corresponding comprehensive score; Among the current model parameters and the updated model parameters, save the model parameters with higher overall scores and determine whether the iteration termination condition is met; If the iteration termination condition is met, the optimal model parameter scheme is output based on the saved model parameters; If the iteration termination condition is not met, the saved model parameters are re-input into the basic simulation simulator as the current model parameters to optimize and update the model parameters until the iteration termination condition is met.

[0015] Furthermore, the adjustment of the initial 3D model based on the optimal model parameter scheme to output the final 3D model includes: When the initial 3D model is a parametric model representation, the values ​​of each model parameter in the optimal model parameter scheme are assigned to the corresponding parameters of the initial 3D model to generate the final 3D model and output it. When the initial 3D model is represented by a patch mesh model, the corresponding key control points of the initial 3D model are displaced according to the displacement values ​​of each key control point in the optimal model parameter scheme. The mesh vertex positions are then updated based on the displaced key control points to generate and output the final 3D model.

[0016] Furthermore, after generating the final 3D model, the following steps are also performed: The final 3D model is physically simulated using a verification simulation simulator, and the model state and structural parameters of the final 3D model under various physical fields are obtained based on the simulation results. The feasibility of the final 3D model is verified based on the model state and structural parameters under each physical field. If the feasibility verification is passed, the final 3D model will be output and visualized. If the feasibility check fails, repeat the iterative search for model parameters.

[0017] A 3D modeling system based on exploratory spatial learning and simulation optimization, used to perform any of the above-mentioned 3D modeling methods, including: The human-computer interaction interface is used to receive user input and visualize the final output 3D model. The initial model building module is used to match or generate an initial 3D model in response to user input requirements. The optimization search module is used to predict and generate the corresponding model parameter exploration space based on the model representation type of the initial 3D model through neural network. The optimization objective is to obtain the highest score output by the simulation simulator according to the preset scoring index. Combined with the preset optimization strategy, the module iteratively searches the model parameters in the exploration space to obtain the optimal model parameter scheme. The model adjustment module is used to adjust the initial 3D model based on the optimal model parameter scheme and output the final 3D model.

[0018] The beneficial effects of this invention are: By automatically matching or generating initial 3D models in response to user needs, and using neural networks to automatically predict the exploration space of corresponding model parameters based on the model representation type, the system can adaptively handle 3D models with different representations, improving modeling adaptability. Furthermore, the exploration space of model parameters generated through neural network prediction covers the globally feasible region of the model parameters, avoiding the exclusion of potential optimal solutions due to experience limitations when manually setting parameter ranges, thus ensuring the accuracy of 3D modeling. Simultaneously, since the exploration space is obtained based on the prediction of the effective parameter distribution, the simulation simulator does not need to perform simulation evaluations on invalid or infeasible parameter combinations during subsequent iterative searches within the exploration space, effectively saving simulation resources and improving overall optimization efficiency. Based on this, automatic iterative optimization is performed with the goal of achieving the highest simulation simulator score, avoiding repetitive manual operations during the modeling process, thereby improving modeling efficiency while ensuring design accuracy in complex modeling scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a process of the present invention; Figure 2 This is a schematic diagram of an exploration space generation process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an iterative search process for model parameters when there is a need for physical simulation, according to an embodiment of the present invention. Figure 4 This is a schematic diagram of an iterative search process for model parameters when there is no physical simulation requirement, according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Example: A 3D modeling method based on exploratory space learning and simulation optimization, such as Figure 1 As shown, it includes: Responding to user input, matching or generating an initial 3D model; Based on the model representation type of the initial 3D model, the exploration space corresponding to the model parameters is generated by predicting through a neural network; The optimization objective is to achieve the highest score output by the simulation simulator based on the preset scoring index. The optimal model parameter scheme is obtained by iteratively searching the model parameters in the exploration space in combination with the preset optimization strategy. The initial 3D model is adjusted based on the optimal model parameter scheme to output the final 3D model.

[0022] The system receives user input in real time via a human-computer interface. This input can be one or more forms of design input, such as text descriptions, sketches, partial 3D models, or others. For text descriptions, keywords can be extracted through semantic parsing to obtain corresponding design parameters, such as dimensions, shape, and functional requirements. For sketches, image recognition or contour extraction is used to convert them into initial geometric features.

[0023] In addition, the human-computer interface also provides preview and editing functions for the initial 3D model, displays intermediate and final results in the current optimization process, and provides options for setting the search process and parameters.

[0024] The obtained design parameters or initial geometric features are then used as search keywords to perform a search and match in a preset 3D model database. If a model that meets the preset similarity threshold exists, it is used as the initial 3D model. If the search results do not meet the threshold, the initial 3D model is directly generated based on the corresponding design parameters or initial geometric features. The preset similarity threshold can be set according to actual needs, but it should be set based on the principle that a higher threshold results in a higher degree of fit with the requirements. In this embodiment, the preset similarity threshold is set to a range of 0.7 to 0.95, with a default value of 0.85, which can balance the matching recall and accuracy.

[0025] It should be noted that the priority of the matching or generation methods can be set according to the actual situation. For example, you can generate first and then match or match first and then generate. When generating first and then matching, the initial 3D model will be generated first. If it cannot be generated, then the matching search will be performed.

[0026] Specifically, an initial 3D model can be generated using a deep generative network.

[0027] The 3D model can be a partially completed 3D model or a 3D model that needs to be matched.

[0028] If the input requirement information is a partial 3D model, it is directly used as the baseline for the initial model.

[0029] For input requirements that require matching a 3D model, the 3D model is used as a constraint, and an initial 3D model is generated by combining it with other requirements.

[0030] Simultaneously, the model further learns the distribution patterns and constraint relationships of model parameters from historical modeling data through neural networks. This allows the model to directly output the variable range or deformable displacement range of each parameter, forming a complete parameter exploration space. This eliminates the need for manual design variables, ranges of variation, and constraints, avoiding improper setting of design variable ranges or omission of constraints due to limitations in human experience. The generated exploration space covers the feasible region of the parameters and reflects the coupling dependencies between variables, thus providing an accurate and efficient search boundary for modeling optimization and ensuring the accuracy of subsequent optimal solution finding.

[0031] Based on this, the model parameters of the initial 3D model are input into the simulation simulator. After physical simulation, a comprehensive score is calculated. Combined with a preset optimization strategy, the model parameters are iteratively optimized with the goal of maximizing the comprehensive score, thereby obtaining the globally optimal solution—the optimal 3D model that meets user requirements. The simulation simulator transforms the trial-and-error iteration of manual optimization into an automatic search, avoiding the randomness of manual adjustments and improving modeling efficiency. Furthermore, the optimization iteration process in the simulation simulator is conducted within the established exploration space, ensuring that all parameter combinations attempted meet basic feasibility constraints, avoiding invalid calculations, and guaranteeing both efficiency and accuracy.

[0032] Finally, the initial 3D model is adjusted based on the optimal model parameter scheme obtained through optimization and iteration to obtain the final 3D model, thereby mapping the optimized model parameters into a visualized solid geometric model to meet user needs.

[0033] In this process, after establishing an initial 3D model based on user-input requirements, it's important to consider that current 3D design employs two model representation types: parametric models and patch mesh models. Parametric models rely on structured dimensions, angles, and other parameters to drive model shape changes, adjusting discrete design parameters. Patch mesh models, on the other hand, rely on the displacement of mesh vertices and key control points to achieve shape deformation, adjusting spatial coordinate points. The optimization objects and transformation rules of these two models differ. Furthermore, during the optimization process of 3D modeling, the parameter value range, control point deformation range, and constraints between variables are mostly set manually based on experience. This not only highly depends on the designer's ability but also increases the computational load of subsequent optimization if the range is too wide, while a range that is too narrow can easily lose the optimal design solution, affecting modeling accuracy.

[0034] Therefore, we further use neural networks to learn historical modeling data offline to obtain the distribution pattern of model parameters, the coupling constraint relationship between parameters and control points, and then combine it with the type of the current initial 3D model to output the corresponding parameter value range or control point deformation displacement range. This constructs an accurate parameter exploration space to fully cover the feasible design area. Combined with the established initial 3D model, it provides an accurate optimization basis for 3D modeling optimization and ensures the accuracy of the final 3D model generated later.

[0035] Specifically, the initial 3D model is represented by either a parametric model or a patch mesh model, and the corresponding exploration space generation process is as follows: Figure 2 As shown.

[0036] For different model representation types of the initial 3D model, the corresponding prediction method is matched, and the model representation type is specifically identified through the data structure of the initial 3D model.

[0037] If the initial 3D model consists of editable geometric constraints and specific model parameters, it can be identified as a parametric model representation. If the initial 3D model consists only of discrete vertex coordinates, patch indices, and topological relationships, it can be identified as a patch mesh model representation.

[0038] Wherein, when the model representation type of the initial 3D model is a parametric model representation, an exploration space for corresponding model parameters is generated based on the model representation type of the initial 3D model through neural network prediction, including: Based on historical modeling data, the parameter sequence distribution and constraint relationship of model parameters are learned through neural networks, and the range of value changes of each model parameter is predicted, forming the exploration space of the corresponding model parameters.

[0039] For an initial 3D model represented by a parametric model, its external dimensions and overall structure are entirely determined by a continuous sequence of parameters. These parameters are high-dimensional correlated variables, and the range of values ​​for a single parameter is jointly restricted by multiple other parameters, forming a complex hierarchical coupling constraint relationship. If parameter intervals or constraint rules are defined manually, the resulting intervals are based solely on subjective experience, and their accuracy cannot be guaranteed. Furthermore, manual analysis of the constraint relationships between such high-dimensional correlated variables is prone to omissions, leading to a large number of invalid solutions in the subsequent optimization process and affecting solution efficiency. Therefore, using historical modeling data as learning samples, a neural network is introduced to learn the parameter sequence distribution characteristics and coupling constraint relationships between parameters, thereby predicting the accurate range of values ​​for each model parameter and forming the corresponding exploration space for the model parameters.

[0040] Specifically, the method of predicting the range of values ​​for each model parameter by learning the parameter sequence distribution and constraint relationships of the model parameters through a neural network based on historical modeling data includes: Numerical parameters of historical parameterized models are extracted from historical modeling data, corresponding parameter sequences are constructed, and parameter sequences of all historical parameterized models are integrated to obtain a training sample set. With the goal of fitting the distribution of parameter sequences in the training sample set, the neural network is trained to learn the parameter sequence distribution and constraint relationship of the model parameters based on the training sample set. The parameter sequence of the initial 3D model is input into the trained neural network. The neural network delineates the basic boundaries of each model parameter value based on the learned parameter sequence distribution, and corrects the basic boundaries of the values ​​in combination with the constraint relationship, thereby obtaining the range of value variation of each model parameter.

[0041] Extract the stored historical modeling data, filter out all historical parametric models, and extract all their numerical parameters in a preset fixed order to form a numerical parameter sequence.

[0042] The extracted numerical parameter sequences are cleaned to remove sequences with missing data and abnormal parameters, and then standardized. The results are then integrated to form a training sample set.

[0043] Based on the Transformer architecture, a deep neural network consisting of multi-head self-attention layers and feedforward network layers is constructed. The parameter sequences in the training dataset are used as input, and the deep neural network is trained to learn the conditional probability distribution of the parameter sequences through autoregression.

[0044] During training, for each parameter sequence, the deep neural network sequentially uses all parameters preceding each parameter in the sequence as input to predict the value of that parameter. Specifically, starting with the first parameter in the sequence, the deep neural network receives a start marker, predicts the value of the first parameter, then receives the actual value of the first parameter, predicts the value of the second parameter, receives the actual values ​​of the first two parameters, predicts the value of the third parameter, and so on, until the entire parameter sequence is predicted. The prediction output at each step is the conditional probability distribution of the corresponding parameter value, and this conditional probability distribution is specifically described by the mean and variance.

[0045] For each parameter in the parameter sequence, the distribution predicted by the deep neural network is compared with the true parameter value, the corresponding negative log-likelihood loss is calculated, and the average of the negative log-likelihood losses of all parameters in the parameter sequence is taken as the total loss of the corresponding parameter sequence.

[0046] The total loss of all training samples is summed and averaged, and the gradient is calculated using the backpropagation algorithm. The network parameters are then updated using the gradient descent optimizer.

[0047] The process of repeating parameter numerical prediction and network parameter updates continues until the loss converges or a preset number of iterations are reached, completing the training of the deep neural network. After training, the acquired deep neural network can output the conditional distribution of the next parameter for any given sequence of preceding parameters, meaning it has learned the dynamic statistical dependencies between parameters. This learned statistical dependency is the constraint relationship of the model parameters. Simultaneously, the deep neural network can calculate the joint likelihood value by multiplying the conditional probabilities of all positions within the parameter sequence to determine the degree to which the set of parameter values ​​conforms to the historical data distribution.

[0048] For the current initial 3D model, the corresponding model parameters are extracted according to the order corresponding to the historical modeling data, forming a model parameter sequence, which is then input into the trained deep neural network. Based on the conditional distribution rules learned during training, the deep neural network outputs a conditional distribution for each parameter in the parameter sequence. Specifically, for the first parameter, the deep neural network outputs its distribution based on the initial label; for the second parameter, it outputs its distribution based on the actual value of the first parameter, and so on.

[0049] Based on the output distribution, the mean values ​​corresponding to each model parameter are determined. Using the mean value as the center and the standard deviation of a preset multiple as the radius, the corresponding upper and lower limits are determined to obtain the basic boundaries. The preset multiple can be set according to actual needs, but it must be set based on the confidence interval of the corresponding normal distribution. In this embodiment, the value range of the preset multiple is set to 2 to 3, with a default value of 2.5, to balance the search efficiency and comprehensiveness of parameter exploration.

[0050] Since the obtained basic boundaries only utilize information from preceding parameters and do not consider global coupling constraints such as the inverse constraints of subsequent parameters on preceding parameters and common constraints among multiple parameters, the boundary correction of each model parameter in the model parameter sequence is further performed using the joint likelihood evaluation capability of a deep neural network.

[0051] Specifically, one model parameter is randomly selected as the target parameter. The actual values ​​of all other parameters in the model parameter sequence, excluding the target parameter, are fixed. Multiple candidate values ​​are uniformly sampled within the basic boundary of the target parameter. For each candidate value, the original parameter value is replaced, and a complete parameter sequence is reconstructed and input into the deep neural network. The deep neural network calculates the joint likelihood of this complete sequence, which is the product of the conditional probabilities at all positions. A higher joint likelihood value indicates that the set of parameters conforms more closely to the statistical regularities and implicit constraints in historical data. Candidate values ​​with joint likelihood values ​​higher than a preset threshold are retained, and the minimum and maximum values ​​of these candidate values ​​are taken as the corrected range of variation for the parameter. The preset threshold for the joint likelihood value can be set according to actual needs, but it must be based on screening parameter combinations that conform to historical statistical regularities and eliminating low-probability invalid solutions. In this embodiment, the preset threshold for the joint likelihood value is set to 0.6 to 0.8 times the average joint likelihood value of the training sample set, with a default value of 0.7 times, to balance parameter feasibility and exploration space.

[0052] The correction operation is performed sequentially on each model parameter in the model parameter sequence. Since the ranges of each model parameter may influence each other, multiple iterations can be performed. Based on the correction range obtained in the previous round, the steps of fixing other parameters, sampling candidate values, calculating joint likelihood, and screening thresholds are repeated until the fluctuation value of the range of all parameters is less than a preset fluctuation threshold or a preset number of rounds is reached, thus determining the value variation range of all model parameters. The preset fluctuation threshold of the parameter range fluctuation value can be set according to actual needs, but it should be set based on the condition that the boundary correction convergence is determined when the fluctuation is below this proportion. In this embodiment, the preset fluctuation threshold of the parameter range fluctuation value corresponds to 1% to 5% of the parameter base boundary width, with a default value of 3%, to balance correction accuracy and computational efficiency.

[0053] Furthermore, based on the range of values ​​for all model parameters, an exploration space for the parameterized model is constructed. This exploration space is the feasible search domain for optimizing model parameters, corresponding to the range of values ​​for each design parameter and the coupling constraints between parameters. Specifically, the exploration space of the parameterized model is a high-dimensional feasible domain that corresponds one-to-one with the dimensions of the model variables. The number of dimensions is consistent with the number of variables in the model design parameters, with each dimension corresponding to the range of values ​​for one variable. The coupling constraints between dimensions limit the combination rules of the variables. Any combination of parameters within the exploration space satisfies basic geometric rationality and historical statistical regularity, and will not encounter fundamental failure problems such as parameter conflicts or mesh distortion. It can be directly used as candidate input for simulation optimization.

[0054] Unlike parametric models, patch mesh models are composed of numerous patches, mesh vertices, and key control points. Adjustments and changes in the model's shape are primarily achieved by altering the spatial positions of these key control points. The key control points in this type of model are not independent; they are constrained by factors such as mesh topology, surface continuity, mechanical deformation, and appearance. There are corresponding linkage constraints between adjacent key control points and global key control points. If the deformable displacement range of control points is set manually based on experience, its accuracy remains difficult to guarantee. Therefore, when the initial 3D model is represented as a patch mesh model, an exploration space for the corresponding model parameters is generated through neural network prediction, including: Based on the initial 3D model, key control points are extracted, and combined with historical modeling data, the deformation displacement range of each key control point is predicted through a neural network, forming an exploration space for the corresponding model parameters.

[0055] Furthermore, using historical modeling data as learning samples, the deformation distribution patterns of different key control points and the topological constraints and linkage deformation rules between control points are learned through neural networks, thereby predicting the displacement range of each key control point during deformation, so as to form a corresponding parameter exploration space.

[0056] Specifically, the process of extracting key control points based on the initial 3D model and combining them with historical modeling data to predict the deformation displacement range of each key control point through a neural network, thereby forming an exploration space for the corresponding model parameters, includes: Extract the key control points of the grid and the corresponding historical deformation data from the historical modeling data, and construct a deformation training sample set; With the goal of fitting the deformation distribution of key control points in the mesh in the deformation training sample set, a neural network is trained based on the training sample set to learn the deformation distribution law and deformation constraint relationship of the key control points in the mesh. Key control points are extracted from the initial 3D model. These control points are then input into the trained neural network. The neural network delineates the basic boundaries of deformation for each control point based on the learned deformation distribution rules. It also corrects the basic boundaries of deformation by combining deformation constraint relationships, thereby obtaining the deformation displacement range of each key control point.

[0057] The stored historical modeling data is extracted, and all historical mesh models are selected. For each historical mesh model, a corresponding bounding mesh is generated, and the coordinates of all key control points of the bounding mesh are extracted. Simultaneously, the actual geometric shape of each historical mesh model relative to a preset reference shape, such as the initial undeformed state of the model or the displacement of a standard template, is used as the deformation displacement value of the corresponding key control point. The coordinates of all key control points of the historical mesh model and their corresponding historical displacement values ​​constitute a deformation sample. Furthermore, based on the generated bounding mesh, the fixed connection relationships between each key control point are determined for use in the subsequent edge structure definition of the graph neural network.

[0058] The extracted deformation samples are cleaned to remove samples with missing or abnormal displacement data. The coordinates and displacement values ​​of key control points are standardized and then integrated to form a deformation training sample set.

[0059] Based on the graph neural network architecture, a neural network consisting of multiple stacked graph convolutional layers and fully connected layers is constructed. The deformation samples in the deformation training sample set and the fixed connection relationships between the determined key control points are used as input. Through multi-layer graph convolution operations, the features of each key control point and its neighboring nodes are aggregated in each layer to learn the joint probability distribution of the displacement of the key control points and the displacement cooperative constraints between adjacent key control points.

[0060] During training, for each deformation sample, the node features of all its key control points and the neighborhood connectivity obtained according to the fixed edge structure are used as input. After multi-layer graph convolution, the statistical distribution parameters of the displacement of each key control point are output, specifically the mean and standard deviation.

[0061] With the training objective of minimizing the negative log-likelihood loss of displacements at all key control points, the network parameters are updated through backpropagation and gradient descent optimizer, enabling the neural network to learn the deformation distribution pattern of each key control point, i.e., the common center location and fluctuation range of displacement. The process of updating and optimizing the network parameters is repeated until the loss converges.

[0062] After training, the coordinates of key control points in the initial 3D model are extracted and input into the neural network. Based on the learned deformation distribution pattern, the neural network outputs the mean and standard deviation of displacement for each key control point. Using the mean as the center and a preset multiple of the standard deviation as the radius, the upper and lower limits of each key control point in the three coordinate directions are determined, forming the basic boundaries of deformation for each key control point. The preset multiple of the standard deviation can be set according to actual needs, but it also needs to be set based on the confidence interval of the corresponding normal distribution. In this embodiment, the preset multiple of the standard deviation ranges from 2 to 3, with a default value of 2.5, to balance the comprehensiveness of deformation exploration with the rationality of mesh deformation.

[0063] Since the basic boundaries obtained are based solely on the independent statistics of each key control point, the collaborative constraints between key control points are not taken into account. For example, the displacement of adjacent key control points should remain smooth and the volume change should be limited. However, during the training process, the displacement prediction of each key control point depends not only on its own features but also on the features of its neighboring key control points. For each key control point, the neural network will perform message passing in the multi-hop neighborhood. This makes the hidden feature vector of the key control point implicitly encode the displacement information of the surrounding key control points, thereby learning the collaborative change law between points, that is, the deformation constraint conditions of the key control points.

[0064] Based on this, the basic boundary is further modified according to the deformation constraint relationship learned by the neural network. During modification, a key control point is used as the target control point, and the displacement values ​​of other key control points are fixed to their current actual values. Multiple candidate displacement values ​​are uniformly sampled within the initial boundary of the target control point. Each candidate value, together with the fixed displacement values ​​of other key control points, forms a complete displacement vector, which is then input into the network to calculate the joint likelihood of this displacement vector. Candidate displacement values ​​with a joint likelihood higher than a preset threshold are retained, and the minimum and maximum values ​​of these candidate values ​​are taken as the modified deformation displacement range of the target control point.

[0065] Corrections are applied sequentially to all key control points, with multiple iterations, until the fluctuation value of the displacement range of each key control point is less than the corresponding preset fluctuation threshold. Ultimately, each key control point obtains a deformation displacement range that satisfies both single-point statistical regularities and global collaborative constraints. The deformation displacement ranges of all key control points collectively constitute the exploration space of the model parameters for the patch mesh model. The preset fluctuation threshold for the displacement range of key control points can be set according to actual needs, but the boundary correction convergence should be determined based on the fluctuation being below this proportion. In this embodiment, the preset fluctuation threshold for the parameter range ranges corresponds to 1% to 5% of the width of the basic displacement range of the control point, with a default value of 3%, to balance correction accuracy and computational efficiency. The exploration space of the model parameters of the patch mesh model is a high-dimensional feasible region that corresponds one-to-one with the dimensions of the model variables. The number of its dimensions is consistent with the number of variables in the key control points of the model. Each dimension corresponds to the range of values ​​of a variable, and the coupling constraints between the dimensions limit the combination rules of the variables. Any combination of parameters in the exploration space satisfies the basic geometric rationality and historical statistical regularity, and there will be no basic failure problems such as parameter conflict and mesh distortion. It can be directly used as a candidate input for simulation optimization.

[0066] After constructing the parameter exploration space, the feasible region for the entire 3D modeling optimization is essentially defined. All parameter values ​​and key control point deformation combinations within this region have undergone dual verification through distribution patterns and coupling constraints, preventing issues such as mesh distortion and parameter conflicts. However, 3D modeling optimization is a typical multi-objective optimization problem, where there are often trade-offs between multiple objectives such as structural performance and material cost. Furthermore, the feasible region contains a massive number of candidate solutions. Relying on manual traversal and trial and error in such a high-dimensional parameter space is clearly inefficient.

[0067] Therefore, based on this, we further select and construct a corresponding simulation simulator and match a suitable solution algorithm according to the differentiability condition of the current simulation environment, so as to ensure the solution efficiency and accuracy of the iterative search while ensuring that it fits the corresponding actual simulation environment.

[0068] Then, taking the highest score output by the corresponding simulation simulator according to the preset scoring index as the optimization target, and combining the preset optimization strategy, the model parameters are iteratively searched in the exploration space to obtain the optimal model parameter scheme.

[0069] The preset scoring indicators specifically include structural mechanics indicators, material consumption indicators, and geometric quality indicators. The structural mechanics indicators specifically include model stress level and structural stability, while the geometric quality indicators include geometric continuity and mesh distortion rate. The model stress level indicator uses the model's maximum stress coefficient and stress uniformity coefficient as defined parameters. The maximum stress coefficient is scored as the difference between 1 and the actual maximum stress or the material's allowable stress. The stress uniformity coefficient is scored as the ratio of the model's average stress to the maximum stress. The average of these two values ​​is taken as the final score for this indicator. The structural stability index is defined by structural safety factor, motion interference compliance, and motion range matching degree. The structural safety factor is calculated as the ratio of the allowable stress of the material to the actual maximum stress. The score is linearly mapped to the range of 0-1 based on the safety factor, with a full score awarded when the safety factor is not less than 2. Motion interference compliance and motion range matching degree are calculated based on the mechanism's motion range and interference conditions. When there is no motion interference, interference compliance receives a full score; when interference exists, it decreases linearly with increasing interference. Motion range matching degree is the ratio of the actual achievable motion range to the design requirement range; the closer the ratio is to 1, the higher the score. The geometric continuity index is defined by the curvature smoothing coefficient, which can be calculated from the curvature distribution of the model. The score is the difference between 1 and the normalized deviation of the curvature change. The mesh distortion rate index is defined by the percentage of qualified mesh elements, with the ratio of qualified mesh elements to the total number of mesh elements directly used as the index score. The material consumption index is defined by the proportion of material volume, calculated from the model's volume statistics. The score is the difference between 1 and the model's solid volume or minimum bounding box volume; lower material usage results in a higher score. The weights of each evaluation index can be set according to actual needs. In this embodiment, the weight of the structural mechanics index is set to 0.5, the weight of the material consumption index is 0.3, and the weight of the geometric quality index is 0.2, prioritizing structural performance while considering material cost and model quality. The preset optimization strategy specifically includes: Based on the initial 3D model, determine whether the current modeling satisfies the differentiability condition; If satisfied, a differentiable physical simulation simulator is constructed, and a gradient descent optimization strategy is matched. The optimization objective is to maximize the score output by the physical simulation simulator based on the preset scoring index, and the model parameters are iteratively searched in the exploration space. If the conditions are not met, a basic simulation simulator is constructed and a reinforcement learning optimization strategy is matched. The optimization objective is to maximize the score output by the basic simulation simulator based on the preset scoring index, and the model parameters are iteratively searched in the exploration space.

[0070] In the automatic optimization process of 3D modeling, the construction type of the simulation simulator depends on whether the physical simulation link defined by the initial 3D model is differentiable. The differentiability is specifically determined by the mathematical differentiability of the constraint functions in the initial 3D model.

[0071] If the geometric constraints involved in the initial 3D model, such as the constraint functions corresponding to stretching, bending, and collision, are all continuously differentiable functions, then the gradients of these functions exist and can be calculated analytically. This makes the energy potential function differentiable with respect to the position of the mesh vertex, and the entire forward computation chain from input parameters to output score is differentiable in principle.

[0072] In such cases, it is determined that the current simulation satisfies the differentiability condition, and a corresponding simulation simulator with differentiable properties is constructed. Gradient information is calculated using automatic differentiation or adjoint methods to fine-tune the model parameters with high precision, thereby efficiently converging to the optimal solution. Specifically, a differentiable physical simulation simulator is used as the simulation simulator that satisfies the differentiability condition. Using this differentiable physical simulation simulator, the entire process of solving physical dynamics, updating model states, and calculating multi-index scores is constructed as a differentiable computational link. Through differentiation operations, the partial derivatives of the comprehensive score with respect to each model parameter or control point displacement, i.e., the gradient information, are calculated. This determines the magnitude of the impact of parameter changes on the comprehensive score and the optimization direction, improving the efficiency and accuracy of optimization modeling.

[0073] like Figure 3 As shown, the construction of a differentiable physical simulation simulator, coupled with a gradient descent optimization strategy, aims to achieve the highest possible score output by the physical simulation simulator based on a preset scoring metric. The process involves iteratively searching the model parameters within the exploration space, including: Obtain or generate a mesh representation of the initial 3D model, define geometric constraints and energy potential functions based on the mesh representation, and construct a differentiable physical simulation simulator; A comprehensive scoring function is constructed based on preset scoring indicators and their weights, and a loss function for iterative search is constructed based on the comprehensive scoring function. Input the current model parameters of the initial 3D model into the physical simulation simulator, and perform forward simulation to obtain the model state; Based on the model state obtained from the simulation, the comprehensive score and the corresponding loss function value are calculated, and the gradient of the loss function with respect to the current model parameters is calculated based on the differential calculation and the loss function value. Based on the calculated gradient, the model parameters are updated using a gradient descent optimization strategy. Among the current model parameters and the updated model parameters, save the model parameters with higher overall scores and determine whether the iteration termination condition is met; If the iteration termination condition is met, the optimal model parameter scheme is output based on the saved model parameters; If the iteration termination condition is not met, the saved model parameters are re-input into the physical simulation simulator as the current model parameters to optimize and update the model parameters until the iteration termination condition is met.

[0074] The initial 3D model is discretized into a mesh representation, that is, the parametric model is processed into triangular facets, or an existing facet mesh model is directly used as a discrete mesh. Based on this mesh representation, geometric constraint functions are defined, including triangular constraints for simulating in-plane tension and shear, dihedral bending constraints for simulating out-of-plane bending, and collision constraints for preventing self-penetration.

[0075] All constraints are integrated into a constraint function, which, combined with the inverse compliance matrix, forms the energy potential function, the expression of which is: ; in, The value of the energy potential function. The position of the grid vertex, To integrate the constraint functions of all constraints, It is the inverse compliance matrix. This is the matrix transpose.

[0076] Specifically, the XPBD (Extended Position Based Dynamics) model is used as the dynamic solution basis for the physical simulation simulator. The gradient of the energy potential function with respect to position is transformed into the acceleration of the mesh vertices, thereby driving the update of vertex positions and changes in velocity. The corresponding dynamic expression is as follows: ; in, For the quality matrix, The velocity of the grid vertices, The position of the grid vertex. Indicates the intermediate prediction state. To predict speed, Momentum is used to describe the amount of motion of a mesh vertex after it has been subjected to an external force, before it has been corrected by geometric constraints. The value of the energy potential function. The constraint force is virtual and is a force generated by the constraint definition. It describes the restoring force that the geometric constraint exerts on the mesh vertices, and its magnitude is proportional to the degree of constraint violation.

[0077] Based on the formed energy potential function and the corresponding XPBD model, a physical simulation simulator is constructed. This physical simulation simulator can receive model parameters and perform forward physical simulation. All mathematical operations during the construction process are in differentiable form, and the simulation link of the physical simulation simulator has the corresponding differentiable capability.

[0078] Then, a comprehensive scoring function is constructed based on the preset scoring indicators and corresponding weights, and the negative value of the comprehensive score is used as the loss function. Specifically, the comprehensive score is calculated by normalizing the scores of each item and weighted summing. Each indicator is uniformly converted to a single score in the range of 0 to 1, and then weighted according to the corresponding weights. The weighted indicators are then summed to obtain the final comprehensive score. The comprehensive score ranges from 0 to 1, and the higher the value, the better the overall performance of the model.

[0079] The model parameters of the initial 3D model are input into a differentiable physics simulation simulator for forward physics simulation. During simulation, the preset simulation time step is... The total number of simulation time steps is .

[0080] The model parameters are converted into the initial positions and velocities of the mesh vertices, and in each time step, they are derived from the current state. Moving forward, the position update is solved using a Gaussian-Seidel-like iterative solver, yielding the updated mesh vertex positions, expressed as: ; in, For the updated mesh vertex positions, This represents the position of the current grid vertex. To update the obtained position, To preset the simulation time step, The velocity of the current grid vertex. For the quality matrix, Preset external load.

[0081] Repeat the time stepping process until the total number of simulation time steps is reached, and output the state sequence recorded at all time steps to obtain the complete physical state sequence.

[0082] After completing the forward simulation and obtaining the complete physical state sequence, the corresponding preset evaluation index values ​​are calculated, and the corresponding comprehensive score is obtained through the comprehensive scoring function. The negative value of the comprehensive score is used as the objective function value to be optimized. Based on the differentiable property of the physical simulation simulator, by introducing the adjoint state sequence, the partial derivative of the objective function with respect to the state sequence is mapped to the gradient with respect to the model parameters, avoiding the computational burden caused by differentiating large-scale physical state sequences. The adjoint state is obtained by inversely solving the adjoint equation, which is constructed from the constraint information and state data saved during the forward simulation.

[0083] The gradient is calculated as follows: ; in, Let the objective function value be the one to be optimized. For model parameters, As an adjoint state sequence, derived from the physical state sequence The accompanying states at each time step constitute the structure. For Jacobian matrices, To obtain from model parameters to physical state sequence The forward mapping function is the forward simulation process of the entire physical simulation simulator.

[0084] After obtaining the gradient, a gradient descent optimization strategy is employed to adjust the model parameters in the reverse direction of the gradient according to a preset learning rate, resulting in updated model parameters. A comprehensive score is then calculated for both the pre- and post-update parameter sets, and the set with the higher score is selected as the current optimal model parameter scheme.

[0085] The iteration then checks whether the termination conditions are met, such as whether the preset maximum number of iterations has been reached, whether the objective function value no longer decreases after multiple consecutive iterations, or whether the search time exceeds a preset upper limit. If the termination conditions are met, the currently saved optimal model parameter scheme is output as the final design result. If not, the current optimal model parameters are used as new input, and the forward simulation, gradient calculation, and parameter update are re-executed. This process is repeated until the termination conditions are met.

[0086] The optimized 3D model through physical simulation can be verified in advance to ensure that the 3D model conforms to the motion law of the mechanism, dynamic mechanical requirements and multi-physics constraints, thereby ensuring the reliability and accuracy of the final output 3D model.

[0087] If the constraint function contains non-differentiable operations, such as step judgments, discrete branches, or discontinuous functions, the energy potential function is not differentiable, and a complete differentiable computation link cannot be established. Therefore, the judgment does not satisfy the differentiability condition.

[0088] At this point, a simulation simulator without differentiability is constructed and matched with a reinforcement learning optimization strategy. Through the balance mechanism of exploration and utilization, the parameter combinations are quickly traversed in the feasible space, and converged to a better solution with a small computational cost.

[0089] Among them, the basic simulation simulator only retains the static forward simulation capability, and completes the quantitative scoring of model performance, geometric quality and other dimensions according to the preset scoring index. The calculation speed is relatively fast. Reinforcement learning is a heuristic search algorithm that does not rely on derivative and gradient information. It can directly use the comprehensive score output by the basic simulation simulator as the reward signal to explore within the exploration space of model parameters. By continuously accumulating reward feedback, it gradually optimizes the search direction to quickly converge to the parameter scheme with the optimal comprehensive score.

[0090] Specifically, such as Figure 4 As shown, the construction of a basic simulation simulator and the matching of a reinforcement learning optimization strategy, with the goal of achieving the highest possible score output by the basic simulation simulator based on a preset scoring metric, involves iteratively searching the model parameters within the exploration space, including: Construct a basic simulation simulator and build a comprehensive scoring function based on preset scoring indicators and their weights; The model parameters in the exploration space are encoded as state vectors, and the adjustment of model parameters is defined as actions. The comprehensive score is used as the reward, and a predefined reinforcement learning optimization strategy is used. The current model parameters of the initial 3D model are input into the basic simulation simulator for physical simulation, and the corresponding comprehensive score is calculated by combining the comprehensive scoring function. The action selection is optimized through reinforcement learning. Adjust the current model parameters based on the selected action, and obtain the updated model parameters; Re-enter the updated model parameters into the basic simulation simulator and obtain the corresponding comprehensive score; Among the current model parameters and the updated model parameters, save the model parameters with higher overall scores and determine whether the iteration termination condition is met; If the iteration termination condition is met, the optimal model parameter scheme is output based on the saved model parameters; If the iteration termination condition is not met, the saved model parameters are re-input into the basic simulation simulator as the current model parameters to optimize and update the model parameters until the iteration termination condition is met.

[0091] A basic simulation simulator is constructed, which includes a finite element analysis module, a kinematics check module, a geometry rendering module, and a scoring index calculation module. The scoring index calculation module is configured with specific calculation methods for preset scoring indices.

[0092] Then, based on the corresponding indicator weights, all preset scoring indicators are weighted and summed to construct a comprehensive scoring function.

[0093] The model parameters within the exploration space are encoded as state vectors. For example, all parameter values ​​of the parameterized model are concatenated into a vector in a fixed order, or the coordinates of all control points of the mesh model are flattened into a vector. The adjustment operations of the model parameters are defined as the action space, with each action corresponding to an incremental adjustment of one parameter or a coordinated displacement of a set of control points.

[0094] A multilayer perceptron-structured policy network is constructed. This network takes a state vector as input and the probability distribution of each action in the action space as output. A proximal policy optimization algorithm is used as the policy gradient update method. Hyperparameters such as the learning rate are set, and the comprehensive score output by a basic simulation simulator is used as the reward signal. Based on the policy network, the policy gradient update method, and the reward signal, a predefined reinforcement learning optimization policy is completed. The training objective of the policy network is to maximize the cumulative discount reward, that is, to continuously adjust the network parameters through learning so that the network's output action probability distribution tends to select actions that yield high rewards.

[0095] The current model parameters of the initial 3D model are input into the basic simulation simulator for physical simulation. The basic simulation simulator first updates its internal 3D geometric model based on the input model parameters, and then sequentially calls the finite element analysis module to calculate the stress distribution and maximum stress value of the model under a given load, calls the kinematics check module to calculate the motion range and interference of the mechanism, and calls the geometry rendering module to generate a visual image of the model and extract features such as curvature distribution. Based on these calculation results, the corresponding scoring index values ​​are calculated sequentially according to the preset scoring index calculation method.

[0096] The current comprehensive score is calculated based on the preset scoring index values ​​using a comprehensive scoring function.

[0097] The current model parameters of the initial 3D model are then encoded into an initial state vector and input into the policy network. The policy network outputs an action probability distribution and selects an action from this distribution using a probability sampling method. Specifically, a uniform random number between 0 and 1 is first generated, and then the cumulative probability of each action is calculated. When the cumulative probability is greater than the random number for the first time, the action is selected.

[0098] Adjust the current model parameters based on the selected action to obtain updated model parameters. Re-input the updated model parameters into the basic simulation simulator to calculate the updated comprehensive score. Compare the current comprehensive score with the updated comprehensive score, and save the one with the higher score as the current optimal model parameter scheme.

[0099] Simultaneously, experience samples are compiled based on the current state, selected action, obtained reward, and updated state, and stored in the experience replay buffer. Every predetermined number of steps, a batch of samples is randomly sampled from the experience replay buffer. The loss function of the policy network is calculated using a proximal policy optimization algorithm, and the parameters of the policy network are updated through backpropagation and gradient descent, making the policy network more inclined to select actions that can obtain higher rewards.

[0100] The system determines whether the iteration termination condition is met, such as reaching a preset maximum number of iterations, the overall score no longer improving after multiple consecutive rounds, or exceeding a preset search time. The preset maximum number of iterations can be set according to the model complexity; the lower the model complexity, the smaller the preset maximum number of iterations. The determination that the overall score no longer improves after multiple consecutive rounds is based on the absence of effective improvement in the overall score for a preset number of consecutive rounds. Effective improvement is specifically the relative improvement in the overall score, calculated by dividing the improvement in the current round's optimal overall score by the optimal overall score of the previous round. If the improvement in the optimal overall score for multiple consecutive rounds is less than a preset convergence threshold, then no improvement is determined. This threshold can be set according to actual needs; in this embodiment, it is specifically set to 0.05% to 0.5%. The preset search time can also be set according to actual needs, specifically selected based on hardware computing power. If one of the iteration termination conditions is met, such as reaching the preset maximum number of iterations, the overall score no longer improving after multiple consecutive rounds, or the total optimization iteration time exceeding the preset search time, then the reinforcement learning iteration is considered converged, and the saved optimal model parameter scheme is output as the final design result. If none of the iteration termination conditions are met, the current optimal model parameters are used as the new current model parameters, and the model parameter update and optimization process is repeated until the termination conditions are met.

[0101] After obtaining the optimal model parameter scheme, the initial 3D model is further adjusted based on the optimal model parameter scheme to output the final 3D model, including: When the initial 3D model is a parametric model representation, the values ​​of each model parameter in the optimal model parameter scheme are assigned to the corresponding parameters of the initial 3D model to generate the final 3D model and output it. When the initial 3D model is represented by a patch mesh model, the corresponding key control points of the initial 3D model are displaced according to the displacement values ​​of each key control point in the optimal model parameter scheme. The mesh vertex positions are then updated based on the displaced key control points to generate and output the final 3D model.

[0102] If the initial 3D model is a parametric model representation, the optimal model parameter scheme contains the final value of each model parameter, such as stretch length, rotation angle, wall thickness, and hole diameter. These parameter values ​​are assigned one by one to the corresponding parameter items in the initial 3D model according to their parameter names or parameter identifiers. The geometry of the model is then recalculated based on the new parameter values, thereby automatically generating an updated 3D geometric model.

[0103] If the initial 3D model is represented by a patch mesh model, the optimal model parameter scheme includes the final displacement value of each key control point, that is, the movement distance in the three coordinate axes. First, extract all key control points of the initial 3D model. For each key control point, calculate the new coordinate position based on its corresponding displacement value, where the new coordinate is equal to the original coordinate plus the displacement value.

[0104] After all critical control points have been updated, the CAGE deformation algorithm or the barycentric coordinate interpolation method is used to update the vertex coordinates of the entire mesh model according to the new positions of the critical control points. Specifically, for each vertex in the patch mesh model, the displacement is calculated based on its weight relationship with the critical control points to obtain the new vertex position after deformation. After all vertices are updated, the final 3D mesh model is generated.

[0105] To ensure the reliability of the final 3D model, the following steps are performed after generating the final 3D model: The final 3D model is physically simulated using a verification simulation simulator, and the model state and structural parameters of the final 3D model under various physical fields are obtained based on the simulation results. The feasibility of the final 3D model is verified based on the model state and structural parameters under each physical field. If the feasibility verification is passed, the final 3D model will be output and visualized. If the feasibility check fails, repeat the iterative search for model parameters.

[0106] The verification simulation simulator differs from the simulation simulator used in the search iteration. It specifically includes modules such as finite element analysis, fluid dynamics calculation, thermodynamic analysis, and manufacturing feasibility assessment, which can evaluate the performance of the three-dimensional model under preset working conditions.

[0107] First, a mesh suitable for finite element analysis is automatically generated based on the final 3D model representation. For parametric models, a volume mesh is generated through triangulation; for patch mesh models, mesh quality optimization and volume mesh conversion are performed directly. The generated mesh includes node and element information and is assigned material properties.

[0108] Simulations of various physical fields were performed sequentially using a calibration simulation simulator. In structural mechanics analysis, boundary conditions and loads were set, and the equilibrium equations were solved using the finite element method to obtain the displacement, strain, and stress components of each node. The maximum stress value, deformation displacement, and safety factor were then calculated. In fluid dynamics analysis, a fluid domain mesh was established, and boundary conditions such as inlet velocity and outlet pressure were set. The Navier-Stokes equations were solved to obtain the velocity field, pressure field, and flow resistance. In thermodynamics analysis, a heat source, convection boundary, and initial temperature were set, and the heat conduction equations were solved to obtain the temperature distribution field and heat transfer coefficient. In manufacturing feasibility analysis, the minimum wall thickness was calculated based on the model's geometric features, the overhang angle was identified, and the need for a supporting structure was determined. Material utilization was also calculated.

[0109] After all simulation calculations are completed, verification indicators are extracted from the simulation results, including the model state under each physical field, such as stress distribution cloud map, fluid trace and temperature cloud map, as well as numerical parameters, such as maximum stress value, minimum safety factor, maximum deformation, pressure loss, highest temperature and minimum wall thickness.

[0110] Based on the model state and structural parameters, the feasibility of the final 3D model is verified by a preset judgment rule, which can be set according to actual needs.

[0111] If all preset judgment rules are met, the feasibility check can be passed, and the final 3D model can be output and displayed in 3D on the visualization interface of the human-computer interaction interface.

[0112] If any preset judgment rule is not met, the feasibility check fails, triggering a new iterative search. Based on the optimal model parameter scheme saved in the previous optimization process, the model parameter iterative search is repeated until a final 3D model that passes the feasibility check is obtained, ensuring the reliability of the final 3D model. The final 3D model displayed for visualization is the optimal model that meets the user's needs.

[0113] Another aspect of this embodiment provides a 3D modeling system based on exploratory space learning and simulation optimization, including: The human-computer interaction interface is used to receive user input and visualize the final output 3D model. The initial model building module is used to match or generate an initial 3D model in response to user input requirements. The optimization search module is used to predict and generate the corresponding model parameter exploration space based on the model representation type of the initial 3D model through neural network. The optimization objective is to obtain the highest score output by the simulation simulator according to the preset scoring index. Combined with the preset optimization strategy, the module iteratively searches the model parameters in the exploration space to obtain the optimal model parameter scheme. The model adjustment module is used to adjust the initial 3D model based on the optimal model parameter scheme and output the final 3D model.

[0114] The human-computer interface (HCI) receives user input and visualizes the final output 3D model. Users can input design requirements through various methods, including text descriptions, hand-drawn sketches, reference images, or by directly uploading partially completed 3D models. The HCI provides a graphical front-end with text input boxes, a sketchpad, a file upload button, and a 3D model preview window. The front-end packages the user input data and sends it to the back-end processing engine. Once the final 3D model is generated, the HCI can display the model through a 3D rendering engine, supporting interactive operations such as rotation, scaling, and sectioning, while also providing a model export interface.

[0115] The backend processing engine specifically includes an initial model building module, an optimization search module, and a model adjustment module. This is achieved using data processing equipment such as a computer with corresponding data processing capabilities. The optimization search module is connected to the initial model building module, and the model adjustment module is connected to both the initial model building module and the optimization search module. Furthermore, the human-computer interaction interface is connected to each of these modules, allowing for data extraction and parameter setting operations.

[0116] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.

Claims

1. A 3D modeling method based on exploratory space learning and simulation optimization, characterized in that, include: Responding to user input, matching or generating an initial 3D model; Based on the model representation type of the initial 3D model, the exploration space corresponding to the model parameters is generated by predicting through a neural network; The optimization objective is to achieve the highest score output by the simulation simulator based on the preset scoring index. The optimal model parameter scheme is obtained by iteratively searching the model parameters in the exploration space in combination with the preset optimization strategy. The initial 3D model is adjusted based on the optimal model parameter scheme to output the final 3D model.

2. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 1, characterized in that, The initial 3D model is represented by either a parametric model representation or a patch mesh model representation.

3. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 2, characterized in that, When the model representation type of the initial 3D model is a parametric model representation, the exploration space for generating the corresponding model parameters is predicted by a neural network, including: Based on historical modeling data, the parameter sequence distribution and constraint relationship of model parameters are learned through neural networks, and the range of value changes of each model parameter is predicted, forming the exploration space of the corresponding model parameters.

4. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 3, characterized in that, The method of predicting the range of values ​​for each model parameter based on historical modeling data, by learning the parameter sequence distribution and constraint relationships of the model parameters through a neural network, includes: Numerical parameters of historical parameterized models are extracted from historical modeling data, corresponding parameter sequences are constructed, and parameter sequences of all historical parameterized models are integrated to obtain a training sample set. With the goal of fitting the distribution of parameter sequences in the training sample set, the neural network is trained to learn the parameter sequence distribution and constraint relationship of the model parameters based on the training sample set. The parameter sequence of the initial 3D model is input into the trained neural network. The neural network delineates the basic boundaries of each model parameter value based on the learned parameter sequence distribution, and corrects the basic boundaries of the values ​​in combination with the constraint relationship, thereby obtaining the range of value variation of each model parameter.

5. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 2, characterized in that, When the initial 3D model is represented as a patch mesh model, an exploration space for predicting and generating corresponding model parameters is generated through a neural network, including: Based on the initial 3D model, key control points are extracted, and combined with historical modeling data, the deformation displacement range of each key control point is predicted through a neural network, forming an exploration space for the corresponding model parameters.

6. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 5, characterized in that, The process involves extracting key control points based on an initial 3D model, combining this with historical modeling data, and using a neural network to predict the deformation displacement range of each key control point, thus forming an exploration space for the corresponding model parameters. This includes: Extract the key control points of the grid and the corresponding historical deformation data from the historical modeling data, and construct a deformation training sample set; With the goal of fitting the deformation distribution of key control points in the mesh in the deformation training sample set, a neural network is trained based on the training sample set to learn the deformation distribution law and deformation constraint relationship of the key control points in the mesh. Key control points are extracted from the initial 3D model and input into the trained neural network. The neural network delineates the basic boundary of deformation for each key control point based on the learned deformation distribution law, and corrects the basic boundary of deformation in combination with deformation constraint relationship to obtain the deformation displacement range of each key control point.

7. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 1, characterized in that, The preset optimization strategy includes: Based on the initial 3D model, determine whether the current modeling satisfies the differentiability condition; If satisfied, a differentiable physical simulation simulator is constructed, and a gradient descent optimization strategy is matched. The optimization objective is to maximize the score output by the physical simulation simulator based on the preset scoring index, and the model parameters are iteratively searched in the exploration space. If the conditions are not met, a basic simulation simulator is constructed and a reinforcement learning optimization strategy is matched. The optimization objective is to maximize the score output by the basic simulation simulator based on the preset scoring index, and the model parameters are iteratively searched in the exploration space.

8. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 7, characterized in that, The construction of a differentiable physical simulation simulator, coupled with a gradient descent optimization strategy, aims to achieve the highest possible score output by the simulator based on a preset scoring metric. The process involves iteratively searching the model parameters within the exploration space, including: Obtain or generate a mesh representation of the initial 3D model, define geometric constraints and energy potential functions based on the mesh representation, and construct a differentiable physical simulation simulator; A comprehensive scoring function is constructed based on preset scoring indicators and their weights, and a loss function for iterative search is constructed based on the comprehensive scoring function. Input the current model parameters of the initial 3D model into the physical simulation simulator, and perform forward simulation to obtain the model state; Based on the model state obtained from the simulation, the comprehensive score and the corresponding loss function value are calculated, and the gradient of the loss function with respect to the current model parameters is calculated based on the differential calculation and the loss function value. Based on the calculated gradient, the model parameters are updated using a gradient descent optimization strategy. Among the current model parameters and the updated model parameters, save the model parameters with higher overall scores and determine whether the iteration termination condition is met; If the iteration termination condition is met, the optimal model parameter scheme is output based on the saved model parameters; If the iteration termination condition is not met, the saved model parameters are re-input into the physical simulation simulator as the current model parameters to optimize and update the model parameters until the iteration termination condition is met.

9. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 7, characterized in that, The construction of a basic simulation simulator and matching it with a reinforcement learning optimization strategy, with the goal of achieving the highest possible score output by the basic simulation simulator based on a preset scoring metric, involves iteratively searching the model parameters within the exploration space, including: Construct a basic simulation simulator and build a comprehensive scoring function based on preset scoring indicators and their weights; The model parameters in the exploration space are encoded as state vectors, and the adjustment of model parameters is defined as actions. The comprehensive score is used as the reward, and a predefined reinforcement learning optimization strategy is used. The current model parameters of the initial 3D model are input into the basic simulation simulator for physical simulation, and the corresponding comprehensive score is calculated by combining the comprehensive scoring function. The action selection is optimized through reinforcement learning. Adjust the current model parameters based on the selected action, and obtain the updated model parameters; Re-enter the updated model parameters into the basic simulation simulator and obtain the corresponding comprehensive score; Among the current model parameters and the updated model parameters, save the model parameters with higher overall scores and determine whether the iteration termination condition is met; If the iteration termination condition is met, the optimal model parameter scheme is output based on the saved model parameters; If the iteration termination condition is not met, the saved model parameters are re-input into the basic simulation simulator as the current model parameters to optimize and update the model parameters until the iteration termination condition is met.

10. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 1, characterized in that, The process of adjusting the initial 3D model based on the optimal model parameter scheme to output the final 3D model includes: When the initial 3D model is a parametric model representation, the values ​​of each model parameter in the optimal model parameter scheme are assigned to the corresponding parameters of the initial 3D model to generate the final 3D model and output it. When the initial 3D model is represented by a patch mesh model, the corresponding key control points of the initial 3D model are displaced according to the displacement values ​​of each key control point in the optimal model parameter scheme. The mesh vertex positions are then updated based on the displaced key control points to generate and output the final 3D model.

11. The 3D modeling method based on exploratory space learning and simulation optimization according to claim 10, characterized in that, After generating the final 3D model, the following is also performed: The final 3D model is physically simulated using a verification simulation simulator, and the model state and structural parameters of the final 3D model under various physical fields are obtained based on the simulation results. The feasibility of the final 3D model is verified based on the model state and structural parameters under each physical field. If the feasibility verification is passed, the final 3D model will be output and visualized. If the feasibility check fails, repeat the iterative search for model parameters.

12. A 3D modeling system based on exploratory space learning and simulation optimization, used to execute the 3D modeling method according to any one of claims 1 to 11, characterized in that, include: The human-computer interaction interface is used to receive user input and visualize the final output 3D model. The initial model building module is used to match or generate an initial 3D model in response to user input requirements. The optimization search module is used to generate the exploration space of the corresponding model parameters based on the model representation type of the initial 3D model through neural network prediction. The optimization objective is to obtain the highest score output by the simulation simulator according to the preset scoring index. Combined with the preset optimization strategy, the module iteratively searches the model parameters in the exploration space to obtain the optimal model parameter scheme. The model adjustment module is used to adjust the initial 3D model based on the optimal model parameter scheme and output the final 3D model.

Images