A performance-driven reverse intelligent design system for three-dimensional woven composites

Abstract

The application discloses a three-dimensional woven composite material performance-driven reverse intelligent design system and relates to the field of composite material reverse design.The system comprises a parameterized modeling engine, an automatic simulation process manager, and a proxy model factory.The parameterized modeling engine is used for receiving a process parameter vector and generating a corresponding representative volume unit geometric model.The automatic simulation process manager is used for scheduling an external finite element solver to perform whole-process analysis on the geometric model and extract macro mechanical performance response.The proxy model factory is used for constructing a Kriging proxy model capable of simultaneously outputting predicted mean and predicted variance.A multi-objective optimization solver is used for taking the proxy model as a fast performance evaluator, solving a multi-objective optimization problem with multiple mechanical performances as targets, and outputting a set of Pareto optimal solutions.The application can significantly shorten the design cycle of a complex material system under the premise of ensuring design reliability and global optimality, thereby improving the mining depth of material performance potential and design efficiency.

Description

Technical Field

[0001] This invention relates to the field of reverse design of composite materials, and in particular to a reverse intelligent design system driven by the performance of three-dimensional woven composite materials. Background Technology

[0002] Three-dimensional woven composite materials, due to the presence of reinforcing fibers in the thickness direction, fundamentally overcome the inherent defects of traditional laminated composite materials, such as weak interlaminar strength and easy delamination. They exhibit superior performance in impact resistance, fatigue resistance, and damage tolerance, making them key candidate materials for high-end equipment fields such as aerospace and deep-sea exploration. The macroscopic mechanical properties of these materials are entirely determined by their internal microstructure, which in turn is controlled by a series of process parameters such as warp density, weft density, and yarn specifications. However, the core requirement of engineering design is essentially inverse: starting from the preset macroscopic mechanical performance target, it is necessary to deduce the optimal combination of process parameters to achieve that performance. The key challenge in achieving this inverse design lies in the complex characteristics of this mapping chain, which involves multi-parameter coupling, strong nonlinearity, and multi-scale correlations, making it impossible to directly establish an inverse function from performance to process using traditional mechanical formulas.

[0003] Current methods for addressing these problems mainly fall into two categories, both of which suffer from significant technical bottlenecks. The first category is trial-and-error methods based on experience or simplified models. This method heavily relies on expert knowledge for manual parameter adjustments and iterations, resulting in long design cycles, enormous resource consumption, and unavoidable design quality instability. Furthermore, it cannot conduct systematic multi-objective optimization exploration within the global design space, easily overlooking solutions with superior performance potential. The second category is simulation-driven optimization, which directly couples finite element simulation with optimization algorithms. While it can provide more accurate evaluations based on physical mechanisms, it requires calling computationally expensive, detailed finite element models in each iteration, leading to single design evaluations taking hours or even days, with total optimization time often reaching months – unacceptable in engineering practice. In addition, although recent research has introduced surrogate model techniques to accelerate evaluation, these generally suffer from shortcomings such as blind initial sampling leading to insufficient information, a lack of quantification and self-calibration capabilities for predictive uncertainty, and a failure to deeply integrate surrogate model construction with multi-objective optimization problems. These shortcomings often cause the global optimization process to fall into pseudo-optimal or violate design constraints, failing to provide robust and reliable reverse design solutions. Therefore, based on the above-mentioned challenges, this invention proposes a reverse intelligent design system driven by the performance of three-dimensional woven composite materials. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a reverse intelligent design system driven by the performance of three-dimensional woven composite materials. This system fundamentally transforms the material design paradigm from an experience-based trial-and-error approach to a data- and knowledge-driven on-demand customization model. While ensuring design reliability and global optimization, it shortens the design cycle of complex material systems from months to days, thereby significantly improving the depth of material performance potential exploration and design efficiency.

[0005] To achieve the above objectives, this invention provides a reverse intelligent design system driven by the performance of three-dimensional woven composite materials. The technical solution is as follows: First, a parametric modeling engine automatically generates a unit cell geometric model containing weaveability constraints based on process parameters. Then, an automated simulation process manager schedules a finite element solver to perform high-throughput simulations, obtaining macroscopic mechanical performance responses to build a sample library. Based on this, a surrogate model factory constructs a Kriging surrogate model capable of simultaneously outputting predicted mean and variance, and introduces an active learning mechanism based on expectation-based improvement of the acquisition function. This mechanism sequentially adds sample points in an uncertainty-driven manner, achieving adaptive convergence of the model's global accuracy with minimal simulation iterations. Next, a multi-objective optimization solver uses the high-precision surrogate model as a fast performance evaluator, employing a multi-objective evolutionary algorithm based on a reference point mechanism for global optimization, outputting a Pareto-optimal process solution set representing the trade-offs between performance parameters. Finally, a visual decision support platform performs high-fidelity verification of candidate solutions and assists users in making multi-criteria decisions through interactive multi-dimensional views, outputting the final design scheme.

[0006] In a first aspect, the present invention provides a reverse intelligent design system driven by the properties of three-dimensional woven composite materials, comprising: The parametric modeling engine receives process parameter vectors and generates representative volumetric unit geometric models that are non-penetrating and satisfy weaveability constraints, based on internally encapsulated mesostructure topology mapping rules containing interference self-correction mechanisms. An automated simulation process manager is used to automatically schedule an external finite element solver to perform a full-process analysis of the geometric model and extract the macroscopic mechanical performance response based on a preset analysis template containing multiple load conditions. The surrogate model factory is used to construct a Kriging surrogate model that can simultaneously output the predicted mean and the predicted variance. The model takes process parameters as input and mechanical properties as output. The surrogate model factory also constructs a data acquisition function based on the predicted variance to drive an adaptive active learning loop to improve the global prediction accuracy of the model in the entire design space with the fewest simulation iterations. A multi-objective optimization solver is used to use the surrogate model as a fast performance evaluator to solve a multi-objective optimization problem with multiple mechanical properties as objectives within the process feasible domain, and outputs a set of Pareto optimal solutions that characterize the optimal trade-offs between performances. A visual decision support platform is used to perform high-fidelity simulation verification of candidate solutions in the Pareto optimal solution set, and to display the mapping relationship between process parameters and post-verification performance of each candidate solution through interactive multi-dimensional view tools, so as to assist users in making a final decision among multiple optimal trade-off solutions.

[0007] Furthermore, the interference self-correction mechanism encapsulated within the parametric modeling engine is implemented as follows: First, the penetration area between yarn entities is detected based on the discrete symbolic distance field. Then, the centerline elastic relaxation iterative algorithm is started to correct the yarn path that interferes until the penetration depth meets the preset tolerance constraint.

[0008] Furthermore, when generating the geometric model, the parametric modeling engine calls the internally integrated weaveability quantification criteria in real time to perform constraint verification on the process parameters. The weaveability quantification criteria include at least: a tight packing criterion based on the maximum fill factor, a curvature damage criterion based on the fiber strain limit, and an interlacing locking criterion based on the buckling amplitude ratio.

[0009] Furthermore, the analysis template encapsulated within the automated simulation process manager contains at least two independent load cases, each load case being matched with a set of periodic boundary conditions adapted to the geometric features of the representative volume element; the automated simulation process manager automatically drives an external finite element solver to sequentially or in parallel execute the solution and result extraction of each load case through a scripted interface, thereby acquiring multidimensional macroscopic mechanical performance responses in batches within a single task scheduling.

[0010] Furthermore, the Kriging surrogate model constructed by the surrogate model factory defines the relationship between process parameters and a single mechanical performance target as a stochastic process consisting of a deterministic global trend term and a local deviation term following a zero-mean Gaussian process. The correlation structure of this stochastic process is defined by an anisotropic kernel function, and its hyperparameters are estimated by maximizing the log-likelihood function under the observed data. By simultaneously providing predicted values ​​and prediction uncertainties, it provides analytical quantitative basis for subsequent active learning sampling decisions, thereby improving the sample efficiency and generalization ability of the surrogate model.

[0011] Furthermore, the acquisition function constructed by the surrogate model factory is the expected improvement function. This acquisition function quantitatively evaluates the expected benefits of adding simulation samples at a point by analytically calculating the expected magnitude by which the performance prediction value at the process parameter point exceeds the current observed optimal value. This drives the global optimization of the newly added sample points, realizing the directional allocation of simulation resources to high-value information regions. Thus, with very few expensive finite element simulations, the surrogate model can achieve high-precision approximation in the global design space.

[0012] Furthermore, the multi-objective evolutionary algorithm run by the multi-objective optimization solver introduces a structured reference point mechanism in environment selection. This mechanism associates individuals in the population with a set of uniformly distributed reference vectors in the target space and performs optimization screening based on the density of individuals associated with each reference vector. This ensures the uniformity and diversity of the distribution of the obtained Pareto front solution set in the multi-dimensional performance target space, thereby providing designers with a comprehensive and homogeneous set of candidate process schemes that characterize the optimal performance trade-off.

[0013] Furthermore, after the visualization decision support platform selects a candidate solution from the Pareto optimal solution set, it triggers the automated simulation process manager to perform high-fidelity finite element verification on the candidate solution with a discretization accuracy higher than that of the mesh used for training the surrogate model, and calculates the relative deviation between the high-fidelity result and the prediction result of the surrogate model, so as to eliminate the risks brought about by the approximation of the surrogate model, ensure that the actual performance of the final recommended solution is highly consistent with the predicted value, and guarantee the success rate of the design.

[0014] Furthermore, the interactive multidimensional view tool in the visualization decision support platform is a parallel coordinate graph, which represents each candidate solution as a broken line spanning multiple parallel coordinate axes. The coordinate axes display process parameters, surrogate model prediction performance, high-fidelity verification performance, and quantitative indicators of constraint satisfaction in parallel. Users can filter, highlight, and compare solutions in a fine-grained manner by interacting with the graph.

[0015] Secondly, a reverse intelligent design method driven by the performance of three-dimensional woven composite materials is also provided. This method is based on the system described in the first aspect and includes: The performance objective inverse design problem is formalized by defining process parameter design variables, multi-objective mechanical performance functions, and constraints. An initial process parameter sample set is generated by space filling experimental design, and automated parametric modeling and finite element simulation are performed on each sample point to obtain the macroscopic performance response in order to build an initial sample library. Based on the sample library, a Kriging proxy model that can provide a measure of predictive variance is trained, and an active learning process driven by uncertainty is executed iteratively: the acquisition function is constructed using the predictive variance to sequentially identify the most informative new process parameter points, the sample library is updated and the model is reconstructed after simulation, until the global prediction accuracy of the model meets the preset criteria. By embedding a high-precision surrogate model into a multi-objective evolutionary algorithm framework, global optimization is performed within the process feasible domain to obtain a set of process parameter schemes composed of Pareto optimal solutions, which characterize the optimal trade-off relationship between multiple mechanical properties. Representative schemes are selected from the Pareto optimal solution set and subjected to finite element verification with a precision higher than that of the surrogate model training. The process parameters, verification performance and target matching degree of each candidate scheme are compared in an interactive visualization environment to help select the final engineering implementation scheme.

[0016] This invention proposes a performance-driven reverse intelligent design system for three-dimensional woven composite materials, fundamentally solving the long-standing bottleneck problems in the design field of three-dimensional woven composite materials, namely the difficulty of reverse solution and the low efficiency of forward design caused by multi-parameter coupling, strong nonlinearity, and multi-scale correlation in the process-performance mapping chain. By systematically integrating surrogate model technology, active learning strategies, and multi-objective optimization algorithms, a leap in design paradigm is achieved from experience-dependent iterative trial and error to data- and knowledge-driven intelligent optimization. On the one hand, by approximating simulation evaluation through surrogate models and combining an uncertainty-guided efficient sample acquisition strategy, the overall computation time and simulation resource consumption required to complete complex multi-objective optimization are reduced by orders of magnitude while ensuring global design accuracy, greatly improving design efficiency. On the other hand, this method can systematically explore the high-dimensional process design space, discover and present globally optimal performance trade-offs that transcend traditional experience, and ensure that the obtained design scheme has performance limits, engineering manufacturability, and result reliability through built-in weaveability constraints and high-fidelity verification, thus providing solid technical support for the high-performance, short-cycle, and customized development of high-end composite material components.

[0017] Beneficial effects By implementing the reverse intelligent design system for three-dimensional woven composite materials driven by the performance of the present invention, the following technical effects are achieved: (1) By combining discrete symbol distance field detection with a centerline elastic relaxation iterative algorithm, automatic detection and correction of geometric interference between yarn entities is achieved. Simultaneously, a multi-dimensional quantitative criterion for weaveability, encompassing tight packing limits, fiber curvature damage thresholds, and interlacing locking conditions, is integrated. This method fundamentally solves the problem in traditional parametric modeling where mesh generation fails due to yarn penetration, rendering the model unusable for finite element analysis. It also ensures that unmanufacturable process parameter combinations can be filtered out in real-time during the automatic generation of the geometric model, thus constructing a zero-intervention, highly reliable geometric front-end for subsequent high-throughput simulation and global optimization.

[0018] (2) Utilizing the inherent ability of the Kriging model to quantify prediction uncertainty, an analytical form of the expected improved acquisition function was constructed to quantitatively evaluate the simulation value of each point in the design space, thereby driving the sequential supplementation of samples. This mechanism breaks through the limitations of blind initial sampling and low utilization of sample information in traditional surrogate model construction, realizing the targeted allocation of expensive simulation resources to high-value information regions. This enables the surrogate model to achieve high-precision approximation in the global design space with the fewest finite element simulations, thus significantly reducing the time and computational cost required for accurate modeling.

[0019] (3) A high-precision Kriging surrogate model that has converged through active learning training is embedded into a multi-objective evolutionary algorithm. A structured reference point mechanism is introduced in the environment selection, and the optimal selection is performed by associating individuals of the population with reference vectors that are uniformly distributed in the target space. On the one hand, this framework replaces finite element analysis with the near-instantaneous evaluation speed of the surrogate model, which solves the bottleneck of limited call count in simulation-driven optimization. On the other hand, through the diversity preservation strategy guided by reference points, it ensures that the obtained Pareto front solution set is uniformly and completely distributed in the multidimensional performance target space, avoiding the defects of traditional methods that are prone to getting trapped in local optima or solution set clustering, and providing designers with a comprehensive and homogeneous set of global optimal trade-offs.

[0020] (4) A progressive design verification process was constructed, from global optimization of the surrogate model to high-fidelity finite element verification. For candidate schemes selected from the Pareto solution set, simulation verification with higher mesh accuracy was performed to eliminate the error risk caused by the approximation of the surrogate model. High-dimensional process parameters and verification performance data were visualized side-by-side through interactive parallel coordinate plots. While retaining the acceleration advantages of the surrogate model, this mechanism ensures the real performance reliability of the final recommended scheme through an independent high-fidelity verification process. It also assists designers in efficiently balancing multi-objective conflicts and understanding parameter sensitivity relationships with an intuitive multi-dimensional view, thereby significantly improving the quality of multi-criteria decision-making and the engineering confidence of the design finalization. Attached Figure Description

[0021] To make the above-described reverse intelligent design system for three-dimensional woven composite material performance driven by the present invention more apparent and understandable, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the method described in this application; Figure 2 This represents a block diagram illustrating the reverse engineering optimization of the system architecture. Figure 3 This is a visual example used for comparing multiple options in decision-making. Detailed Implementation

[0023] Example 1: This embodiment uses a three-dimensional woven composite material design example to illustrate the technology, and the method flow is as follows: Figure 1 As shown.

[0024] The design goal is to achieve the following while strictly controlling the fiber volume fraction to (54% ± 2%): and It has high stiffness in both main load-bearing directions and is as balanced as possible, that is, maximizing the warp modulus and the weft modulus and seeking the best trade-off between the two.

[0025] Define design variables and its domain For composite materials with defined components, key and independent process parameters are taken as design variables and set... ,in and These are the warp and weft densities (threads / cm) of a 2.5D woven structure, respectively. The objective function is determined as follows: The goal is and Maximize. Define constraints, with fiber volume fraction as the constraint. Since parametric models can automatically guarantee geometric weaveability, no additional geometric constraints are required.

[0026] Generated using optimal Latin hypercube sampling initial sample points This ensures that the samples are evenly distributed within the design space, avoiding clustering. Each point in The automated analysis process is executed sequentially: the modeling script is invoked, and input is performed. The system automatically generates a solid geometric model of a 3D woven unit cell based on predefined topological mapping rules. It then calls an analysis script to sequentially import the geometric model into commercial finite element software, assign material properties, construct the material coordinate system, mesh the data, set analysis steps, apply periodic boundary conditions, apply loads, and submit the analysis job. After solving, it calls a results processing script to read the target performance from the results file. , pairing data Store in a structured database.

[0027] Based on this, the surrogate model is constructed, validated, and adaptively enhanced. The sample data is divided into a 95% training set and a 5% validation set. The Kriging surrogate model is defined as follows: Assuming a performance target... (like ) is a design variable The implementation of the stochastic process, namely ,in The global mean. It is a function with a mean of 0 and a covariance of For stationary Gaussian processes, an anisotropic Gaussian kernel function is chosen as the correlation function. ,in is a length scale hyperparameter that controls the smoothness of each input dimension. Given... Observation data ,in yes matrix, yes Vector. The likelihood function of a Gaussian process is... ,in, yes The correlation matrix, its elements , It is a vector of all 1s. To simplify calculations, we maximize the log-likelihood. Given , and The maximum likelihood estimate has an analytical solution. Substitute it back to get information about The condensed log-likelihood function The optimal hyperparameters are obtained by maximizing the function using an optimization algorithm (such as L-BFGS-B). For new points The Kriging model provides the predicted distribution. ,in ,in .

[0028] Calculate the root mean square error and coefficient of determination between the predicted and actual values. If the accuracy criteria are not met (e.g.) If this is the case, then an active learning iterative enhancement process is initiated, using the desired improvement (EI) as the acquisition function. The goal is to maximize... ,make The improvement amount is defined as the current best observation value. .because Follows a Gaussian distribution. Its expected value: ,in , and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. As the next sample point, it is returned to the simulation analysis process. New data is added to the database, and the surrogate model is retrained, iterating until the model meets the accuracy criterion. The reverse design optimization system architecture to implement the above method is as follows: Figure 2 As shown.

[0029] After completing the agent model construction, establish a bi-objective optimization mathematical model: Satisfying the design variables and fiber volume fraction constraints, where These are predictions from the Kriging surrogate model. The solution is obtained using a non-dominated sorting genetic algorithm with reference points (NSGA-III). The algorithm initializes the population size. Target number Generate structured reference points On the unit hyperplane of the target space. For two targets, they can simply be uniformly distributed along each target axis. One point. For simplicity, it can be generated directly. A set of weight vectors uniformly distributed within the range (0,1) and summing to 1. In each main loop, offspring are generated through binary tournament selection (based on Pareto rank and crowding / reference point distance), simulated binary crossover (SBX), and polynomial mutation. Merging parent and child generations yields a population of size 2N. Using the proxy model Quick assessment The goals and constraints of all individuals; for Individuals in the dataset are sorted using a fast non-dominated ranking based on their target values ​​to obtain the frontier hierarchy. New populations are filled in according to the frontier hierarchy. When adding an entire layer This could lead to population overpopulation. At that time, for the critical layer Individual execution based on reference point selection: and After standardizing all individual target values, they are associated with the nearest reference line. The number of individuals associated with each reference point is calculated, and the process iteratively starts from the reference point with the fewest associated individuals. Choose one from the individuals In the middle, until Scale reached The selection criteria can be either minimizing the vertical distance of the individual to the reference line (to increase convergence) or random selection (to maintain diversity). After the optimizer finishes running, it obtains a set of Pareto optimal solutions. Each solution represents a solution in and Different combinations of process parameters are used to achieve different trade-offs.

[0030] From the Pareto frontier of the target space Select Representative schemes were selected, covering the endpoints, inflection points, and central regions of the frontier. High-fidelity verification was then performed on the selected candidate schemes using a more precise finite element analysis process (e.g., reducing the mesh size by 50%) to obtain the verification performance values. Calculate the relative error between the predicted values ​​and those of the surrogate model. In this embodiment, all errors were less than 5%, verifying the reliability of the surrogate model and the optimization results.

[0031] Finally, a parallel coordinate graph containing axes such as longitude density, latitude density, prediction and verification performance, volume fraction, and tradeoff coefficients is used for visual decision-making. For example... Figure 3 As shown, the figure contains 8 vertical axes: (The text abruptly ends here, so the translation stops as well.) , latitude ,predict ,verify ,predict ,verify Volume fraction Trade-off coefficient (defined as) The closer to 1, the more balanced. Scheme A ( ): The broken line is in The shaft is the highest. The axis is the lowest, belonging to "high". "Guided" Option B ): The broken line is in The shaft is the highest. The axis is relatively low, belonging to the "high" category. "Guided" approach. Option C ( ): The broken line is in and Both axes are at the upper-middle level and are close in strength, with a tradeoff coefficient of ≈1.1. Through interactive comparison, scheme C provides the most balanced bidirectional high stiffness while satisfying all constraints. The predictions and verifications showed a high degree of agreement, and these parameters were ultimately selected as the optimal manufacturing parameters. The system automatically generates a design report that includes the optimal process parameters, performance comparisons, uncertainty analysis, and process cards.

[0032] Example 2: This embodiment further illustrates the core algorithm encapsulated within the parametric modeling engine and provides a parametric modeling case with a typical structure to ensure unambiguous mapping from process parameters to high-fidelity geometric models.

[0033] The cross-section of the yarn changes dynamically with position during the weaving process, and the cross-sectional shape is uniformly described using hyperelliptic curves: ,in , Let be the half-width and half-height of the cross section along the path parameter s, respectively. This is the shape index. When... The cross-section is elliptical. The cross-sectional area tends to be rectangular (racetrack-shaped). To achieve precise control of fiber volume content, the cross-sectional area... Constant constraints must be satisfied: ,in For the Gamma function, It is determined by both yarn fineness and fiber volume content. At the interlacing point, the cross-section rotates, and the rotation angle is... It is determined by spherical linear interpolation along the path in the local coordinate system.

[0034] The yarn centerline path is determined by the weave matrix. (indicating the first) root warp and first The interlacing relationship of the root and weft yarns is generated through spline fitting. Five iterations are used. Spline curves construct spatial paths Its expression is ,in Five times spline basis functions This is the control point set. The control point locations are determined by the coordinates of the interlacing points and the thickness compatibility equation. Confirmed, among which For the total thickness of the prefabricated body, this equation ensures spatial interlocking of warp and weft yarns at interlacing points, satisfying the geometric constraints of actual weaving. To eliminate the penetration phenomenon between yarn entities, a hybrid strategy of discrete symbolic distance field plus centerline elastic relaxation is adopted: First, the unit cell is discretized into a voxel grid, and the symbolic distance from each voxel to each yarn surface is calculated. Negative values ​​indicate penetration. If the penetration volume percentage exceeds a threshold (e.g., 0.5%), then centerline iterative relaxation is initiated. ,in This refers to the location of the yarn centerline node. This is the stiffness coefficient. The normal direction is used. Iteration continues until the maximum penetration depth is less than the set tolerance (e.g., 0.01 mm). Finally, the voxels in the penetrated region are assigned to the matrix according to the material allocation rules or weighted by volume fraction.

[0035] The parametric modeling engine calculates the following three types of weaveability quantification criteria in real time as boundary constraints of the feasible region of design variables. The close packing criterion requires a fill factor. ,in and These are the unit cell width and thickness, respectively. Take a value of 0.78–0.85 (the limit for hexagonal close packing). The curvature damage criterion requires the curvature of the yarn centerline. The corresponding strain on the outer side of the fiber It should be less than the fiber fracture strain (1.5%~2.0% for carbon fiber), that is... The interlacing locking criterion requires a buckling amplitude ratio of... Should meet This ensures that the risks of warp breakage and warp sheath opening are controllable. It automatically triggers rejection or correction for parameter combinations that exceed the boundaries.

[0036] The parametric modeling implementation methods for typical structures are described below. The three-dimensional orthogonal structure consists of straight warp yarns, straight weft yarns, and Z-axis binding yarns. The warp yarn path is... Weft path is , The yarn path uses a zigzag pattern: ,in For thickness direction parameters, , For the bundling period, This is the total thickness. The cross-section is assumed to be rectangular. The yarn automatically compresses according to the principle of area conservation when passing through the warp and weft interlacing zone. In the three-dimensional interlocking structure, the warp yarns obliquely pass through the thickness direction at a certain angle to lock multiple layers of straight weft yarns, and the warp yarn path adopts a two-parameter sine curve: ,in For layer index, The interlocking angle is the amplitude that varies with the layer and reaches its maximum value in the middle layer. The angle is controlled between 20° and 35° to ensure weavability. In the three-dimensional interlocking structure, each warp yarn interlocks only with 2 to 3 adjacent weft yarns. The interlocking layers of adjacent warp yarns are staggered, and the path uses segmented splines. ,in , The buckling curves are those within the upper and lower layers. This is a linear transition segment. The misalignment offset sequence is composed of... The three-dimensional knitting structure is generated using a four-step knitting process. The yarn path is generated based on the motion trajectory tracking of the yarn carrier, and the position of the yarn carrier changes with time steps. ,in This is a four-step motion matrix. The yarn space curve is obtained by interpolation from the yarn carrier trajectory: ,in The length of the flower festival. For weaving corners, This represents the unit vector of the azimuth direction. The above process can be implemented through independent programming or secondary development based on fabric modeling software such as Texgen.

Claims

1. A reverse intelligent design system driven by the performance of three-dimensional woven composite materials, characterized in that, include: The parametric modeling engine receives process parameter vectors and generates representative volumetric unit geometric models that are non-penetrating and satisfy weaveability constraints, based on internally encapsulated mesostructure topology mapping rules containing interference self-correction mechanisms. The interference self-correction mechanism encapsulated within the parametric modeling engine is implemented as follows: first, the penetration area between yarn entities is detected based on the discrete symbolic distance field, and then the centerline elastic relaxation iterative algorithm is started to correct the yarn path where interference occurs until the penetration depth meets the preset tolerance constraint. When generating a geometric model, the parametric modeling engine calls the internally integrated weaveability quantification criteria in real time to perform constraint verification on the process parameters. The weaveability quantification criteria include at least: a tight packing criterion based on the maximum fill factor, a curvature damage criterion based on the fiber strain limit, and an interlacing locking criterion based on the buckling amplitude ratio. An automated simulation process manager is used to automatically schedule an external finite element solver to perform a full-process analysis of the geometric model and extract the macroscopic mechanical performance response based on a preset analysis template containing multiple load conditions. The surrogate model factory is used to construct a Kriging surrogate model that can simultaneously output the predicted mean and the predicted variance. The model takes process parameters as input and mechanical properties as output. The surrogate model factory also constructs a data acquisition function based on the predicted variance to drive an adaptive active learning loop to improve the global prediction accuracy of the model in the entire design space with the fewest simulation iterations. A multi-objective optimization solver is used to use the surrogate model as a fast performance evaluator to solve a multi-objective optimization problem with multiple mechanical properties as objectives within the process feasible domain, and outputs a set of Pareto optimal solutions that characterize the optimal trade-offs between performances. A visual decision support platform is used to perform high-fidelity simulation verification of candidate solutions in the Pareto optimal solution set, and to display the mapping relationship between process parameters and post-verification performance of each candidate solution through interactive multi-dimensional view tools, so as to assist users in making a final decision among multiple optimal trade-off solutions.

2. The system according to claim 1, characterized in that: The analysis template encapsulated within the automated simulation process manager contains at least two independent load cases, each load case being matched with a set of periodic boundary conditions adapted to the geometric features of the representative volume element; the automated simulation process manager automatically drives an external finite element solver to sequentially or in parallel execute the solution and result extraction of each load case through a scripting interface.

3. The system according to claim 1, characterized in that: The Kriging surrogate model constructed by the surrogate model factory defines the relationship between process parameters and a single mechanical performance target as a stochastic process consisting of a deterministic global trend term and a local deviation term following a zero-mean Gaussian process. The correlation structure of this stochastic process is defined by an anisotropic kernel function, and its hyperparameters are estimated by maximizing the log-likelihood function under the observed data.

4. The system according to claim 1, characterized in that: The acquisition function constructed by the surrogate model factory is the expected improvement function. This acquisition function quantitatively evaluates the expected benefits of adding simulation samples at a point by analytically calculating the expected magnitude by which the performance prediction value at the process parameter point exceeds the current observed optimal value, and drives the global optimization of the newly added sample points.

5. The system according to claim 1, characterized in that: The multi-objective evolutionary algorithm run by the multi-objective optimization solver introduces a structured reference point mechanism in environment selection. This mechanism associates individuals in the population with a set of uniformly distributed reference vectors in the target space and performs optimization selection based on the density of individuals associated with each reference vector.

6. The system according to claim 1, characterized in that: After selecting a candidate solution from the Pareto optimal solution set, the visualization decision support platform triggers the automated simulation process manager to perform high-fidelity finite element verification on the candidate solution with a discretization accuracy higher than that of the mesh used for training the surrogate model, and calculates the relative deviation between the high-fidelity result and the prediction result of the surrogate model.

7. The system according to claim 1, characterized in that: The interactive multidimensional view tool in the visualization decision support platform is a parallel coordinate graph, which represents each candidate solution as a broken line spanning multiple parallel coordinate axes. The coordinate axes display process parameters, surrogate model prediction performance, high-fidelity verification performance, and quantitative indicators of constraint satisfaction in parallel. Users can filter, highlight, and compare solutions in a fine-grained manner by interacting with the graph.

8. A reverse intelligent design method driven by the properties of three-dimensional woven composite materials, characterized in that: The method is implemented based on the system described in any one of claims 1-7: The method includes: The performance objective inverse design problem is formalized by defining process parameter design variables, multi-objective mechanical performance functions, and constraints. An initial process parameter sample set is generated by space filling experimental design, and automated parametric modeling and finite element simulation are performed on each sample point to obtain the macroscopic performance response in order to build an initial sample library. Based on the sample library, a Kriging proxy model that can provide a measure of predictive variance is trained, and an active learning process driven by uncertainty is executed iteratively: the acquisition function is constructed using the predictive variance to sequentially identify the most informative new process parameter points, the sample library is updated and the model is reconstructed after simulation, until the global prediction accuracy of the model meets the preset criteria. By embedding a high-precision surrogate model into a multi-objective evolutionary algorithm framework, global optimization is performed within the process feasible domain to obtain a set of process parameter schemes composed of Pareto optimal solutions, which characterize the optimal trade-off relationship between multiple mechanical properties. Representative schemes are selected from the Pareto optimal solution set and subjected to finite element verification with a precision higher than that of the surrogate model training. The process parameters, verification performance and target matching degree of each candidate scheme are compared in an interactive visualization environment to help select the final engineering implementation scheme.

Images