A computer-aided design system for a new type of clothing fabric

Abstract

The application discloses a kind of novel computer-aided design system of clothing fabric, and the application relates to the technical field of computer-aided design, system includes fabric parameter mapping module, two-dimensional pattern design module, graph neural network physical agent simulation module, loss calculation unit and manufacturability constraint solver, fabric parameter mapping module is based on few-sample learning and is mapped into standardization simulation parameter with fabric metadata and mechanical test data, graph neural network physical agent simulation module predicts three-dimensional drape grid node coordinates and stress distribution in differentiable manner, loss calculation unit obtains total loss by weighting topographic error, suture edge alignment error and material strain constraint error, closed-loop iterative correction is carried out to two-dimensional pattern control point by gradient back propagation, realize the automatic closed-loop design between three-dimensional simulation effect and two-dimensional pattern, give consideration to visual effect and industrial manufacturability.

Description

Technical Field

[0001] This invention relates to the field of computer-aided design technology, specifically to a novel computer-aided design system for clothing fabrics. Background Technology

[0002] With the development of the textile industry, a large number of new clothing fabrics with special mechanical properties have emerged, such as regenerated fiber fabrics and intelligent temperature control fabrics. In the existing computer-aided design workflow for clothing, there are usually two relatively independent parts: a two-dimensional pattern design module and a three-dimensional physical simulation module. The two-dimensional pattern design module uses parametric curves to model the outline of the cut pieces. Designers can adjust the shape and size of the cut pieces by dragging control points. The three-dimensional physical simulation module is based on the extended position constraint dynamics method or the finite element method to simulate and display the mechanical effects of hanging, collision and wrinkling of the sewn cut pieces on a virtual human body model, so that designers can intuitively evaluate the final garment effect.

[0003] The current mainstream computer-aided design software for clothing generally adopts the following technical approach: Designers draw the outline curves of the garment pieces and set the stitching relationships in the 2D pattern-making module. The system reads the mechanical parameters of the fabric, which are usually obtained by manually inputting experience values ​​or relying on test reports from offline testing equipment. The system assembles the garment pieces onto a 3D virtual human body model according to the stitching relationships, calls the physical simulation engine to iteratively solve the assembled garment model, and outputs 3D simulation results such as drape shape and wrinkle distribution. Designers observe the 3D simulation effect with their naked eyes. If they find that the drape is not up to standard, the silhouette collapses, or the wrinkles are abnormal, they return to the 2D pattern-making module based on their personal experience, manually adjust the relevant control points or curve parameters, and then resubmit the 3D simulation. This process is repeated until the effect is satisfactory. In this approach, there is no automated mathematical connection between 2D pattern modification and 3D simulation evaluation. The connection between the two relies entirely on manual judgment and repeated iterations.

[0004] The existing technical solutions have the following shortcomings: First, since the traditional collision detection and topology mesh calculation process itself does not have continuous differentiability, the 3D simulation results cannot be directly mapped back to the 2D pattern parameters mathematically. The modification of the 2D pattern depends entirely on the designer's experience and repeated experiments, resulting in an extremely long development and pattern-making cycle for new fabrics. Second, the mechanical measurement system of new fabrics is independent of the physical engine parameter system used by commercial computer-aided design software. The measurement equipment is expensive and historical data is scarce, making it difficult for new fabrics to obtain accurate simulation performance in the system. Third, existing automatic pattern generation or correction methods often only pursue the fit between the 3D visual effect and the target shape, failing to fully consider the manufacturability constraints required by industrial production, such as the matching of the seam length of the cut pieces and the stress characteristics of the fabric yarn direction. As a result, even if the generated pattern is visually highly consistent with the target, it may be scrapped in the workshop production because it does not meet the actual manufacturing process requirements, forming a gap between the simulation effect and industrial production. Therefore, we propose a new computer-aided design system for clothing fabrics. Summary of the Invention

[0005] The purpose of this invention is to provide a novel computer-aided design system for clothing fabrics.

[0006] To address the aforementioned shortcomings of existing technologies, the technical problems to be solved by this invention include the following three aspects: First, how to overcome the non-differentiable barriers between two-dimensional pattern design and three-dimensional physical simulation in a clothing computer-aided design system due to discontinuities in topological changes and collision detection, so that the deviation between the three-dimensional draping simulation effect and the design target can be automatically and continuously fed back to the two-dimensional pattern correction process, avoiding designers relying on experience to repeatedly make manual trials and errors. Second, how to solve the problems of data fragmentation, inconsistent standards, and high acquisition costs of mechanical parameters of new fabrics between different measuring devices and different commercial computer-aided design physical engines, so that new fabrics can obtain accurate simulation performance quickly and at low cost. Third, how to simultaneously ensure that the generated results meet the manufacturability constraints required for industrial production, such as equal seam length, reasonable yarn force, and seam tensile strength, during the automatic generation and correction of patterns, avoiding the problem of ideal simulation effects but actual difficulty in sewing and production.

[0007] To address the aforementioned technical problems, this invention provides a novel computer-aided design system for clothing fabrics, comprising a fabric metadata and mechanical test data input interface, a fabric parameter mapping module, a two-dimensional pattern design module, a graph neural network physical proxy simulation module, a target effect comparison and loss calculation unit, a manufacturability constraint solver, and a gradient backpropagation and pattern control point update unit.

[0008] The fabric metadata and mechanical test data input interface receives basic metadata such as fabric composition ratio, density, and thickness, as well as limited mechanical test data. The fabric parameter mapping module, connected to the input interface, includes a few-shot learning network to map the basic metadata and mechanical test data into standardized simulation parameters, enabling low-cost automatic inversion of simulation parameters for novel fabrics. The two-dimensional pattern design module uses B-spline curve control points to parameterize the geometric data of the cut pieces and records the topological data of the seam relationships between the pieces. The graph neural network physical proxy simulation module is connected to both the fabric parameter mapping module and the two-dimensional pattern design module, receiving the geometric data of the cut pieces, the topological data of the seam relationships, and the standardized simulation parameters. Through a differentiable forward computation method using graph structure message passing, it outputs the predicted coordinates of the three-dimensional overhanging mesh nodes and the node stress distribution, thereby transforming traditional... The non-differentiable physical simulation process is replaced by an end-to-end differentiable forward computation process. The target effect comparison and loss calculation unit compares the predicted 3D overhang mesh node coordinates with the pre-set target 3D effect to calculate the 3D shape error. The manufacturability constraint solver calculates the stitching edge alignment error and material strain constraint error based on the predicted 3D overhang mesh node coordinates, and then weights and sums the 3D shape error, stitching edge alignment error, and material strain constraint error to obtain the total loss. The gradient backpropagation and pattern control point update unit backpropagates along the forward computation path of the graph neural network physical proxy simulation module through the automatic differentiation engine, calculates the gradient of the total loss relative to the 2D pattern control points, and iteratively updates the control points according to the gradient until the total loss is less than a preset threshold. The updated 2D pattern control points are then output, thereby realizing the automatic closed-loop correction of the 2D pattern driven by the 3D effect.

[0009] In one embodiment of the present invention, the system further includes a coarse-to-fine adaptive multi-scale scheduling unit. In the early stage of the pattern correction iteration, a low-resolution grid and coarse-grained graph neural network are used for calculation to quickly approximate the general outline of the target shape. After the iteration reaches the preset stage switching condition, it switches to a high-resolution grid and fine-grained calculation to solve the wrinkle details and local constraints in detail, thereby improving the iteration efficiency of pattern correction while ensuring the optimization accuracy.

[0010] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0011] 1. This invention constructs an end-to-end differentiable graph neural network physical proxy simulation module, which enables a continuous and differentiable mathematical mapping relationship between the three-dimensional simulation results and the two-dimensional pattern control points. It can automatically complete the pattern correction by gradient backpropagation, avoiding the problem of designers relying on experience to repeatedly make manual trials and errors. It greatly compresses the traditional pattern making cycle that relies on repeated manual trials and errors into an automatic closed-loop iteration.

[0012] 2. This invention establishes a unified mapping space between fabric metadata, mechanical parameters and simulation parameters, and combines few-shot learning to achieve parameter inversion. The simulation parameters of new fabrics can be obtained with only a small amount of test data, which significantly reduces the cost and cycle of obtaining simulation parameters of new fabrics, while ensuring simulation accuracy.

[0013] 3. This invention introduces seam edge alignment error and material strain constraint error into the loss function of pattern reverse correction, so that the optimization result can naturally meet the industrial sewing requirements while approximating the target visual effect, eliminating the disconnect between simulation results and actual production, and realizing the design effect that simulation results can be used for production.

[0014] 4. By adopting an adaptive multi-scale scheduling strategy from coarse to fine, this invention improves the iterative efficiency of pattern correction while ensuring optimization accuracy, enabling the pattern correction process to meet the real-time interaction requirements of commercial computer-aided design software. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall closed-loop architecture of the system in this embodiment of the invention;

[0016] Figure 2 This is a schematic diagram of the fabric parameter mapping training process in an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of the automatic pattern correction and iteration process in an embodiment of the present invention. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0019] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] System overall architecture:

[0021] The present invention provides a novel computer-aided design system for clothing fabrics, including a fabric metadata and mechanical test data input interface, a fabric parameter mapping module, a two-dimensional pattern design module, a graph neural network physical proxy simulation module, a target effect comparison and loss calculation unit, a manufacturability constraint solver, a gradient backpropagation and pattern control point update unit, and a coarse-to-fine adaptive multi-scale scheduling unit. The modules are interconnected by data interfaces to form a complete closed loop from fabric parameter input to automatic two-dimensional pattern correction output.

[0022] The fabric metadata and mechanical test data input interface is used to receive basic metadata such as the composition ratio, density, and thickness of the fabric, as well as finite mechanical test data obtained through the mechanical performance testing system, such as measured or partially measured values ​​of tensile modulus, flexural stiffness, and shear modulus. The two-dimensional pattern design module parametrically represents the outline of the cut pieces using several B-spline curve control points and records the topological data of the stitching relationship between the cut pieces. Designers can adjust the initial shape and size of the cut pieces by dragging the control points. The graph neural network physical agent simulation module receives the geometric data of the cut pieces and the topological data of the stitching relationship output by the two-dimensional pattern design module, as well as the standardized simulation parameters output by the fabric parameter mapping module. It predicts the coordinates of the three-dimensional overhanging mesh nodes and the stress distribution of the nodes using a differentiable forward computation method. The target effect comparison and loss calculation unit, the manufacturability constraint solver, and the gradient backpropagation and pattern control point update unit work together to form the core closed loop of this system to realize the automatic correction of the two-dimensional pattern driven by the three-dimensional effect.

[0023] Unified mapping method for fabric mechanical parameters:

[0024] The fabric parameter mapping module addresses the inconsistency between the mechanical parameters of multi-source heterogeneous fabrics and the parameter system of the simulation engine. Internally, it contains a multilayer perceptron mapping network composed of multiple fully connected hidden layers. This mapping relationship can be expressed as:

[0025] ;

[0026] In the formula, This represents the standardized simulation parameter vector output, including the tensile stiffness coefficient, bending stiffness coefficient, damping coefficient, and density parameters required for the fabric elasticity model or extended position-constrained dynamics model. This represents the input feature vector, which is composed of basic metadata of the fabric and finite mechanical test data. This represents the mapping function implemented by the multilayer perceptron. This represents the set of learnable weight parameters for all fully connected layers in the multilayer perceptron. In a specific embodiment, the multilayer perceptron mapping network can be configured to include three fully connected hidden layers, with each layer having 128 neurons, and the activation function can be a linear rectified function.

[0027] The training of the fabric parameter mapping module is divided into an offline training phase and an online few-shot fine-tuning phase. In the offline training phase, known metadata and corresponding standardized simulation parameters from the historical fabric database are used to supervise the training of the multilayer perceptron mapping network, resulting in a pre-trained mapping network. In the online few-shot fine-tuning phase, for new fabrics to be integrated into the system, basic metadata provided by the manufacturer is collected, and a very small number of key mechanical index tests are performed to obtain a small number of labeled samples. The weight parameters of the pre-trained mapping network are then fine-tuned according to the following loss function:

[0028] ;

[0029] In the formula, This indicates the limited number of labeled samples available for the new fabric. Indicates the first The input feature vector of each sample, Indicates the first The true standardized simulation parameters corresponding to each sample are obtained through limited measured data or publicly available industry reference tables. The square of the Euclidean distance is used to measure the deviation between the output value of the mapping network and the true value. After fine-tuning, the complete metadata of the new fabric is input into the fine-tuned mapping network, which outputs the standardized simulation parameters of the new fabric as input parameters for the subsequent graph neural network physical agent simulation module. The above method enables the acquisition of relatively accurate standardized simulation parameters for new fabrics that have not appeared in the historical database, such as regenerated fiber fabrics or smart temperature-controlled fabrics, with only a very small number of actual tests. This avoids the high cost and long cycle caused by conducting comprehensive offline testing on every new fabric.

[0030] Differentiable physics proxy simulation based on graph neural networks:

[0031] The graph neural network physics proxy simulation module replaces the traditional non-differentiable 3D physics simulation process with an end-to-end differentiable graph neural network forward computation process. This module represents the clothing mesh as a graph structure, where each node represents a sampling point on the clothing surface. Node features include the point's current 3D coordinates, velocity, and standardized simulation parameters of the fabric. Edges in the graph structure represent the mechanical connections between adjacent nodes, and edge features include the initial edge length and the rate of change of the current edge length. The graph neural network physics proxy simulation module contains at least two layers of message passing structures. A single-layer message passing process can be represented as:

[0032] ;

[0033] In the formula, Indicates the first Layer nodes eigenvectors, Represents a node The set of neighbor nodes in a graph structure Represents a node With nodes The edge feature vectors between them The message function is used to compute neighbor node pairs. The influence quantity is achieved by a multilayer perceptron. This represents the node update function, used to update the node's own features by combining the aggregated neighbor information. It is also implemented by a multilayer perceptron. Indicates the updated number Layer nodes After passing through multiple layers of message passing, the graph neural network physical agent simulation module decodes the final node features into predicted three-dimensional coordinate displacement and node stress estimates through the decoding layer.

[0034] The offline training process of the graph neural network physical proxy simulation module is as follows: Using the finite element method or extended position-constrained dynamics method, large-scale 3D suspension simulation data is generated in batches for different plate types and different combinations of standardized simulation parameters. This includes node coordinates, velocity, and stress distribution at each time step, serving as the ground truth data for training. Each set of simulation data is converted into a graph structure representation. Node features and edge features are initialized according to the aforementioned steps. A graph neural network with a multi-layer message-passing structure is constructed. The error between the predicted node positions and stress distribution and the ground truth data is used as a supervision signal to train the network parameters offline, resulting in a pre-trained network. The completed graph neural network physical proxy model, in the online design phase, converts the current two-dimensional pattern into an initial graph structure after assembly through stitching relationships. The standardized simulation parameters output by the fabric parameter mapping module are used as node features or global features as inputs to the pre-trained graph neural network physical proxy model. This graph neural network physical proxy model continuously outputs the predicted three-dimensional drape shape and node stress distribution through a forward propagation method composed entirely of tensor operations. Since this forward propagation process is continuously differentiable at every point, it achieves an end-to-end differentiable representation of the three-dimensional simulation results for the two-dimensional pattern parameters, providing a mathematical basis for subsequent gradient-based pattern correction.

[0035] Gradient-based two-dimensional pattern closed-loop inverse correction:

[0036] This module is used to achieve automatic closed-loop correction of the 2D pattern driven by 3D effects, while simultaneously satisfying manufacturability constraints. It is the core module of this invention. The total loss function constructed in this invention is:

[0037] ;

[0038] In the formula, This represents the total loss used to drive pattern control point updates. This represents the three-dimensional topography error term. This indicates the seam alignment error term. This represents the material strain constraint error term. , These represent the weighting coefficients of the seam alignment error and the material strain constraint error relative to the shape error. These coefficients are preset to fixed values ​​by the designer based on the specific fabric and style, and are used to balance the relationship between visual effect approximation and the degree of process constraint satisfaction.

[0039] The three-dimensional topography error term is defined as:

[0040] ;

[0041] In the formula, This indicates the total number of 3D mesh nodes participating in the comparison. The graph neural network physical surrogate model predicts the first... The three-dimensional coordinates of each node, This indicates the target 3D effect set by the designer. The three-dimensional coordinates of each corresponding node;

[0042] The seam alignment error term is defined as follows:

[0043] ;

[0044] In the formula, This represents the set of all edge node pairs that need to be stitched and aligned. and These represent the coordinates of two corresponding nodes on a sewing edge in three-dimensional space. This term is used to constrain the assembled cut pieces to achieve geometric equal length and alignment at the sewing edge, ensuring sewing feasibility.

[0045] The material strain constraint error term is defined as follows:

[0046] ;

[0047] In the formula, This represents the total number of triangular faces in the 3D mesh of the garment. The graph neural network physical surrogate model predicts the first... The dependent variable of a triangular facet. This represents the strain threshold of the fabric in the corresponding yarn direction, determined by the fabric standardization simulation parameters. This item is used to penalize local stress concentration areas that exceed the mechanical tolerance of the material, thus avoiding the generation of patterns that would lead to seam tearing or fabric damage in actual production.

[0048] The gradient update process of pattern control points can be represented as:

[0049] ;

[0050] In the formula, Indicates the first The coordinate vector of the control points of the 2D plate B-spline curve in the next iteration. This represents the learner rate of the optimizer. Indicates total loss Relative to control point vector The gradient is calculated by the automatic differentiation engine through backpropagation along the forward propagation path of the graph neural network. Indicates the first The control point coordinate vector updated in the next iteration;

[0051] In the specific implementation process, the system first receives the initial two-dimensional pattern control point data, the topological data of the stitching relationship between the pattern pieces, and the target three-dimensional effect data set by the designer. This target three-dimensional effect data can be the target drape shape manually adjusted by the designer, or a target mesh obtained from a reference image or a reference object scan. Then, based on the current iteration stage of the coarse-to-fine scheduling unit, the system selects the corresponding resolution mesh discretization method to perform assembly conversion on the current two-dimensional pattern, generating graph structure data. This data is then input into a pre-trained graph neural network physical surrogate model, and the predicted three-dimensional mesh node coordinates and stress distribution are calculated forward. Next, the shape error, stitching edge alignment error, and material strain constraint error are calculated according to the aforementioned formulas, and a weighted sum is obtained to obtain the total loss. Finally, the system uses automatic... The differential engine, along the forward propagation path of the graph neural network and the parameterized geometric mapping path of the pattern, backpropagates to calculate the gradient of the total loss relative to the 2D pattern control point vector. Then, it uses a gradient-based iterative optimization algorithm to update the control point vector once. After that, it checks whether the current iteration count has reached the stage switching condition preset by the coarse-to-fine scheduling unit. If it does, it switches the grid resolution and graph neural network computation granularity from coarse to fine, and enters the fine solution stage. Finally, it checks whether the total loss is less than the preset convergence threshold or whether the number of iterations has reached the preset upper limit. If neither is satisfied, it returns to the aforementioned graph structure transformation process to continue iterating. If it is satisfied, it terminates the iteration and outputs the current control point vector as the correction result of the 2D pattern control points, completing the closed-loop automatic correction of the 2D pattern driven by the 3D effect.

[0052] The coarse-to-fine adaptive multi-scale scheduling unit uses low-resolution meshes and coarse-grained graph neural network calculations in the early stages of iteration, exhibiting relatively sparse mesh nodes and a large iteration step size, which is used to quickly approximate the general outline of the target shape. After the iteration reaches the preset stage switching condition, it switches to high-resolution meshes and fine-grained local calculations, exhibiting denser mesh nodes and a smaller iteration step size, which is used to finely solve wrinkle details and local constraints. In this way, while ensuring optimization accuracy, the pattern correction iteration time is compressed to the order of magnitude that meets the real-time interaction requirements of commercial computer-aided design software.

[0053] The distribution of multiple cut pieces and their control points in the two-dimensional pattern and the three-dimensional draping effect after assembly and sewing are marked with arrowed connecting lines indicating the corresponding seam edges. The figure marks the seam edge pairs that need to meet the equal length constraint, the cut piece areas that need to meet the yarn stress constraint, and the seam areas that need to meet the tensile strength constraint, which are used to illustrate the specific role of manufacturability constraints in the pattern control points and the three-dimensional simulation results.

[0054] Alternative implementation methods:

[0055] In another embodiment of the present invention, a differentiable finite element method or a differentiable position constraint solution method can be used to replace the graph neural network physical proxy model. That is, instead of using a pre-trained graph neural network, the key computational operators such as collision detection softening approximation and constraint solution continuous relaxation in the finite element method or extended position constraint dynamics method are directly modified to be differentiable, so that the traditional physical simulation process itself has end-to-end differentiability. The gradient backpropagation of the three-dimensional simulation results to the two-dimensional pattern control points can also be realized. This implementation method may be better than the proxy model method in terms of computational accuracy, but it is usually not as good as the implementation method using the graph neural network physical proxy model in terms of computational efficiency and real-time interactive performance.

[0056] In another embodiment of the present invention, a self-attention mechanism can be introduced on the basis of the message passing structure of the graph neural network to form a hybrid structure of graph neural network and self-attention network as a physical proxy model, which is used to capture the mechanical coupling relationship between distant nodes in the clothing mesh, such as the linkage draping effect between the hem and the neckline, thereby improving the prediction accuracy of the draping effect of complex silhouettes while ensuring the differentiability of forward propagation.

[0057] Those skilled in the art should understand that the above embodiments are merely specific examples of the present invention, used to illustrate the technical solutions of the present invention, and not to limit its scope of protection. All equivalent substitutions and improvements made to the technical solutions of the present invention within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

[0058] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A novel computer-aided design system for clothing fabrics, characterized in that: include: The fabric metadata and mechanical test data input interface is used to receive basic metadata such as the composition ratio, density, and thickness of the fabric, as well as limited mechanical test data. The fabric parameter mapping module is connected to the input interface of the fabric metadata and mechanical test data. The fabric parameter mapping module contains a few-shot learning network, which is used to map the basic metadata and mechanical test data into standardized simulation parameters. The two-dimensional pattern design module is used to represent garment patterns using the geometric data of the cut pieces parametrically represented by B-spline curve control points and the topological data of the seam relationship. The graph neural network physical agent simulation module is connected to the fabric parameter mapping module and the two-dimensional pattern design module, respectively. It is used to receive the cut piece geometric data and sewing relationship topology data output by the two-dimensional pattern design module and the standardized simulation parameters output by the fabric parameter mapping module. The node features in the graph structure include the vertex three-dimensional coordinates, velocity and the standardized simulation parameters of the fabric to which it belongs, and the edge features include the initial edge length and the rate of change of the current edge length. Through the differentiable forward calculation method of graph structure message passing, the predicted three-dimensional overhanging mesh node coordinates and node stress distribution are output. The target effect comparison and loss calculation unit is connected to the graph neural network physical proxy simulation module and is used to compare the predicted three-dimensional overhanging mesh node coordinates with the pre-set target three-dimensional effect and calculate the three-dimensional shape error. A manufacturability constraint solver, connected to the target effect comparison and loss calculation unit, is used to calculate the stitching edge alignment error and material strain constraint error based on the predicted three-dimensional overhanging mesh node coordinates, and to weight and sum the three-dimensional topography error, the stitching edge alignment error and the material strain constraint error to obtain the total loss; The gradient backpropagation and pattern control point update unit, connected to the manufacturability constraint solver, is used to backpropagate along the forward computation path of the graph neural network physical proxy simulation module through the automatic differentiation engine, calculate the gradient of the total loss relative to the B-spline curve control points in the two-dimensional pattern design module, and iteratively update the control points according to the gradient until the total loss is less than a preset threshold, and output the updated two-dimensional pattern control points. It also includes a coarse-to-fine adaptive multi-scale scheduling unit, used to employ low-resolution mesh and coarse-grained graph neural network computation in the early stages of iteration, and to switch the mesh resolution and graph neural network computation granularity to high-resolution mesh and fine-grained computation when the iteration reaches a preset stage switching condition.

2. The computer-aided design system for a novel garment fabric according to claim 1, characterized in that: The fabric parameter mapping module includes a multilayer perceptron mapping network. The multilayer perceptron mapping network is trained offline under supervised supervision using known metadata and corresponding standardized simulation parameters from a historical fabric database to obtain a pre-trained mapping network. For a new fabric, a small number of labeled samples of the new fabric are collected, and the weight parameters of the pre-trained mapping network are fine-tuned with a few samples to obtain a fine-tuned mapping network. The metadata of the new fabric is input into the fine-tuned mapping network, and the standardized simulation parameters of the new fabric are output.

3. The computer-aided design system for a novel garment fabric according to claim 1, characterized in that: The graph neural network physical proxy simulation module represents the clothing mesh as a graph structure. The nodes in the graph structure correspond to the vertices of the clothing mesh. The node features include the three-dimensional coordinates, velocity, and standardized simulation parameters of the fabric to which the vertex belongs. The edges in the graph structure correspond to the mesh edges and seam edges. The edge features include the initial edge length and the rate of change of the current edge length. The graph neural network physical proxy simulation module uses a multi-layer message passing structure. Each layer of nodes updates its own feature representation by aggregating the features of neighboring nodes and edge features. The final node features are decoded into predicted three-dimensional coordinate displacement and node stress estimates by a decoding layer.

4. The computer-aided design system for a novel garment fabric according to claim 1, characterized in that: The seam alignment error is calculated based on the coordinate difference between two corresponding nodes on the seam to be sewn in three-dimensional space, and the material strain constraint error is calculated based on the portion of the predicted strain of the triangular facet in the three-dimensional mesh of the garment that exceeds the strain threshold of the fabric in the corresponding yarn direction.

5. The computer-aided design system for a novel garment fabric according to claim 1, characterized in that: It also includes a coarse-to-fine adaptive multi-scale scheduling unit, which is connected to the two-dimensional pattern design module and the graph neural network physical proxy simulation module, respectively. It is used to use low-resolution grid and coarse-grained graph neural network calculation in the early stage of iteration, and switch the grid resolution and graph neural network calculation granularity to high-resolution grid and fine-grained calculation when the iteration reaches the preset stage switching condition.

6. The computer-aided design system for a novel garment fabric according to claim 1, characterized in that: The gradient backpropagation and pattern control point update unit adopts a gradient-based iterative optimization algorithm, which obtains the updated coordinate vector of the control point by subtracting the product of the learning rate and the gradient from the coordinate vector of the control point in the current iteration.

7. A novel computer-aided design method for clothing fabrics, characterized in that, include: The basic metadata of the fabric and limited mechanical test data are obtained, and the basic metadata and mechanical test data are input into a pre-trained few-shot learning network to obtain standardized simulation parameters. Obtain the geometric data of the cut pieces and the topological data of the seam relationship, which are parameterized by the control points of the B-spline curve; The geometric data of the cut pieces, the topological data of the stitching relationship, and the standardized simulation parameters are converted into graph structure data, input into a pre-trained graph neural network physical agent model, and the predicted three-dimensional overhanging mesh node coordinates and node stress distribution are obtained through the differentiable forward computation method of graph structure message passing. The predicted coordinates of the three-dimensional overhanging mesh nodes are compared with the pre-set target three-dimensional effect to calculate the three-dimensional shape error. Based on the predicted coordinates of the three-dimensional overhanging mesh nodes, calculate the stitching edge alignment error and the material strain constraint error; The total loss is obtained by weighted summing of the three-dimensional topography error, the stitching edge alignment error, and the material strain constraint error. The gradient of the total loss relative to the control points of the B-spline curve is calculated by backpropagating along the forward computation path of the graph neural network physical proxy model using an automatic differentiation engine. The control points of the B-spline curve are iteratively updated according to the gradient until the total loss is less than a preset threshold, and the updated two-dimensional pattern control points are output.

8. The computer-aided design method for a novel clothing fabric according to claim 7, characterized in that: The method for obtaining the pre-trained few-shot learning network includes: using known metadata from a historical fabric database and corresponding standardized simulation parameters to perform offline supervised training on a multilayer perceptron mapping network to obtain a pre-trained mapping network; for a new type of fabric, collecting a small number of labeled samples of the new fabric and fine-tuning the weight parameters of the pre-trained mapping network with few samples to obtain the pre-trained few-shot learning network.

9. The computer-aided design method for a novel clothing fabric according to claim 7, characterized in that: During the iterative update of the B-spline curve control points, in the early stage of the iteration, a low-resolution grid and a coarse-grained graph neural network are used to calculate the predicted coordinates of the three-dimensional overhanging grid nodes. When the iteration reaches the preset stage switching condition, the grid resolution and graph neural network calculation granularity are switched to a high-resolution grid and fine-grained calculation.

10. The computer-aided design method for a novel garment fabric according to claim 7, characterized in that: The graph neural network physical proxy model represents the clothing mesh as a graph structure. The nodes in the graph structure correspond to the vertices of the clothing mesh. The node features include the three-dimensional coordinates, velocity, and standardized simulation parameters of the fabric to which the vertex belongs. The edges in the graph structure correspond to the mesh edges and seam edges. The edge features include the initial edge length and the rate of change of the current edge length. The graph neural network physical proxy model updates the node features through a multi-layer message passing structure and decodes the final node features into predicted three-dimensional coordinate displacement and node stress estimates through a decoding layer.

Images