Warping prediction optimization method for complex carbon fiber composite suspension structure

Abstract

The invention discloses a warping prediction optimization method for a complex carbon fiber composite suspension structure, and the method comprises the steps: S1, reading CAD and laying layers, and tracking a main heat flow along fibers to segment an anisotropic block; s2, rearranging nodes, locally fitting a principal curvature and a control area, extracting characteristics of a fiber angle, a layer thickness, a temperature and warping, and constructing a weighted adjacency graph; s3, retrieving a historical high-fidelity sample, mapping and enhancing the historical high-fidelity sample into a source domain graph, and performing confrontation training by combining an encoder and a domain discriminator to realize node-level alignment; s4, constructing a generator and a discriminator based on the alignment graph, and carrying out alternate confrontation optimization after preheating; s5, constructing reconstruction, confrontation, peak stress and energy constraint multi-target loss; and S6, the freezing generator reasones initial version warping, and evaluates a node risk clustering high-risk sub-domain. According to the method, on the premise that node-level physical conservation is guaranteed, the warping morphology of the carbon fiber composite component with any shape and fiber laying layer combination under the curing working condition can be rapidly and accurately predicted.

Description

Technical Field

[0001] This application relates to the field of digital simulation and intelligent optimization technology for composite material molding processes, and in particular to a method for predicting and optimizing warping of complex carbon fiber composite suspended structures. Background Technology

[0002] In the field of composite material molding, it has become common practice to use numerical simulation to evaluate the residual stress and warpage generated in carbon fiber laminates during the entire heating-holding-cooling process. Current mainstream solutions are mostly based on transient finite element process simulation: first, the CAD surface is unfolded layer by layer to generate a fine three-dimensional solid or thick-shell mesh; then, the thermal conductivity equation, resin curing kinetics, and anisotropic elasto-thermal constitutive model are coupled, and the time-series temperature field and curing degree field are tracked using explicit or implicit solvers. Finally, chemical shrinkage and thermal expansion are superimposed on the structural field to calculate residual deformation. To improve efficiency, some commercial software introduces laminate equivalence, block-based adaptive meshing, and multi-level order reduction techniques. Through thickness integration or volume averaging methods, multi-layer layups are compressed into a macroscopic material card, and combined with parallel solving and multi-threaded contact algorithms, single-piece simulations can be completed within hours. On the other hand, data-driven approaches are also rapidly developing. Researchers establish mapping relationships between warpage measurement data of historical test pieces or small components and process inputs (such as furnace temperature profiles, layup angle sequences, and fixture preload). Commonly used methods include random forests, convolutional neural networks, and graph convolutional network models that abstract mesh nodes into graph structures. Some works attempt to use low-fidelity finite element results as priors and then use machine learning models for residual correction, aiming to maintain local prediction accuracy while ensuring speed. In addition, cross-domain transfer and domain adaptation frameworks have also been introduced into composite material process simulation. A common approach is to register the mesh shapes of different parts to a common template and then align the statistical distribution in the latent space, thereby sharing existing high-fidelity samples.

[0003] While finite element method (FEM) simulations can provide node-level physical details, when dealing with large-sized components with drastic curvature changes and complex fiber orientations, the mesh size and time step often increase exponentially, resulting in simulation cycles of tens of hours, which is incompatible with the pace of iterative design. Further simplification to thick shells or equivalent laminates leads to the loss of crucial information such as edge peeling and local stress concentration. Purely data-driven methods are sensitive to sample distribution; extrapolation instability easily occurs when the layup angle combination or furnace temperature curve deviates from the training set. Furthermore, the lack of explicit thermo-elastic coupling conservation constraints results in insufficient interpretability and physical consistency. Existing hybrid order reduction and template registration strategies rely on manually defined shape alignment and feature scaling rules in cross-domain scenarios. If the fixture layout or main heat flow path of the target part differs significantly from the source domain, alignment errors are amplified, leading to accumulated warp prediction bias. More importantly, most publicly available solutions lack dynamic refinement mechanisms for high-risk local areas during curing, failing to capture sharp warp changes at curvature peaks or densely packed support points while maintaining overall efficiency.

[0004] The core technical problem that this invention aims to solve is: how to quickly and accurately predict the warping morphology of carbon fiber composite components with arbitrary shapes and fiber layup combinations under curing conditions while ensuring node-level physical conservation, and how to adaptively identify and refine local high-risk areas so that the prediction results meet both engineering timeliness and local accuracy and anisotropic physical consistency. Summary of the Invention

[0005] One objective of this invention is to propose a warpage prediction and optimization method for complex carbon fiber composite suspended structures, which can quickly and accurately predict the warpage morphology of carbon fiber composite components with arbitrary shapes and fiber layup combinations under curing conditions, while ensuring nodal-level physical conservation.

[0006] A method for predicting and optimizing warpage of complex carbon fiber composite suspended structures according to an embodiment of the present invention includes the following steps:

[0007] S1: Read in CAD and layup, trace the main heat flow along the fiber to divide anisotropic blocks, generate equivalent material cards, perform adaptive mesh steady-state solution and obtain warpage prior field;

[0008] S2: Rearrange nodes, locally fit principal curvature and control area, extract fiber angle, layer thickness, temperature, and warpage features, and construct a weighted adjacency graph;

[0009] S3: Retrieve historical high-fidelity samples, map and enhance them into source domain graphs, and conduct adversarial training between the joint encoder and the domain discriminator to achieve node-level alignment;

[0010] S4: Based on the alignment map, a generator and discriminator are constructed, and after preheating, alternating adversarial optimization is performed, with dynamic weighted stress and smoothing of errors until convergence;

[0011] S5: Construct a multi-objective loss mechanism including reconstruction, adversarial, peak stress, and energy constraints; dynamically normalize the gradient and iterate to tolerance by adjusting the weights.

[0012] S6: Freeze the initial warping of the generator inference, evaluate node risk clustering of high-risk subdomains, refine the local grid and iterate incrementally until the model is tolerated and archived.

[0013] Optionally, S1 includes the following steps:

[0014] S11. Read the CAD geometry file and the corresponding layup description file, and perform main heat flow path tracing according to the fiber principal direction and the continuous direction of the outer curvature; when the tracing path encounters a dense area of ​​support points or a transition position of the first derivative of the outer curvature, the tracing is terminated immediately, thereby obtaining a preliminary set of functional blocks that maintain the independence of anisotropic heat conduction. .

[0015] S12. For each preliminary functional block, perform a spatial overlap determination between its boundary surface and the fixture contact line. If the same fixture contact line is found to span multiple blocks, perform geometric Boolean operations on the spanning blocks to parallelize them into a new single block. Alternatively, the clamp contact lines can be re-cut until all clamp contact lines are completely within a single block, ensuring that the boundary conditions for subsequent low-fidelity solutions are closed.

[0016] S13, Block Read fiber layup angle sequence Interlayer thickness sequence and resin type By querying the preset material comparison matrix, the in-plane equivalent thermal expansion coefficient of the block can be obtained. Equivalent elastic modulus equivalent density and specific heat capacity And package it into a simplified material card This reduces the dimensionality of subsequent solution parameters.

[0017] S14, in the block Internally based on the predicted temperature gradient Generate a 3D adaptive mesh. Mesh nodes. Size Calculated using the following formula:

[0018]

[0019] in, As a global baseline length scale; This is the gradient adjustment coefficient; For nodes Temperature gradient prediction model at the location; For nodes The grid size at that location.

[0020] Statistics of all The corresponding total number of nodes, and the set upper limit. Compare; if the total exceeds Then according to The incremental method increases synchronously. The value is maintained until the number of nodes meets the upper limit requirement.

[0021] S15, Block Input furnace temperature curve and materials card The nodal temperatures were obtained by solving the problem. Combined with curing shrinkage ratio With node temperature difference Calculate nodal thermal strain .in This is a reference temperature.

[0022] S16. Apply thermal strain to the node along the thickness direction. Integrate to obtain the spontaneous warping of the block. :

[0023]

[0024] For nodes The initial value of the warping displacement; For blocks The equivalent bending stiffness factor; For blocks The total thickness; For thickness position node Thermal strain.

[0025] The integral results are mapped back to the coarse-grained grid nodes through linear interpolation to form the initial value field for block warping. .

[0026] S17. Perform weighted Laplace smoothing and normalization on the warped initial values ​​of all blocks to remove discontinuities in cross-block values; then, set the node coordinates... Temperature field With normalized warp field Encapsulated as a priori package .

[0027] Optionally, S2 includes the following steps:

[0028] S21. Read the node coordinate set given in step S1. Temperature field With warped field Sort by functional block number first, then within each functional block, use coordinates... of - Component order, - The component order is rearranged in two stages to ensure that neighboring nodes remain continuous in the index space, eliminating the geometric coherence disruption caused by random grid generation.

[0029] S22, at each node At this point, the triaxial principal curvature is calculated using a quadratic surface least squares fitting method based on the set of its one-ring neighborhood nodes. , , Simultaneously, the nodal control area is obtained using the Voronoi area subdivision method. The curvature-area pair jointly describes the degree of abrupt changes in local stiffness, providing a geometric basis for determining warp sensitivity.

[0030] S23, Adjust node normals Projected onto the block fiber principal vector The fiber layup angle at the node is obtained from the included projection angle. Combined with the total interlayer thickness ,Will , and Write node properties to explicitly encode the anisotropic stress response.

[0031] S24. Synchronously load the temperature rise curve of the corresponding node. Record peak temperature and the slope of linear regression during the cooling phase Following the aforementioned geometric and material properties, the complete feature vector of a node is defined as follows:

[0032]

[0033] in, , , For nodes exist , , Principal curvature in the direction; Control area for nodes based on Voronoi area allocation; For the fiber layup angle at the node; The thickness of the local laminate at the node; This refers to the peak temperature at the node. The temperature slope during the node cooling phase; The initial value of node warping is calculated in step S1.

[0034] S25. Within the same functional block, for each node Select a value whose Euclidean distance does not exceed the cutoff radius. All neighboring nodes Forming a connecting edge;

[0035] Combining the principal direction vector of nodal heat flow , With temperature difference Calculate the local edge weights:

[0036]

[0037] in, , The coordinates of the node; The distance attenuation scale; This is the gain coefficient in the heat flow direction; Connect the nodes with unit vectors; This is the temperature difference gain coefficient; , This represents the principal direction vector of heat flow at the node; The temperature difference between nodes; The radius of the cutoff point is the radius of the connecting edge.

[0038] S26. In cross-functional block node pairs First, check the temperature difference. With warp gradient ;when or Cross-block connections are established only when the weights are calculated based on Equation S25, in order to highlight potential warping jump points.

[0039] S27. For all node feature vectors Apply zero-mean unit variance normalization; for all edge weights Layered caching by functional blocks allows subsequent networks to distinguish between local and cross-block coupling strength. (Geometric-physical feature map) Construction complete, including For a set of nodes, For weighted edge set, For node feature matrix, It is a sparse edge weight matrix.

[0040] Optionally, S3 includes the following steps:

[0041] S31. Retrieve high-fidelity thermo-elastic coupling simulation files of completed carbon fiber suspension components from the enterprise database, and search for geometric reference dimensions respectively. Layer sequence With furnace temperature curve Calculate the feature vector corresponding to the target part. Euclidean distance; only retain distances not exceeding a threshold. The samples are used to construct the initial pool of the source domain. Reduce irrelevant structural noise interference during alignment training.

[0042] S32, For the initial pool of the source domain For each simulation record, extract the node temperature. With nodal strain Based on the node index rearrangement table and edge weight rules in step S2, a point-by-point mapping is generated to... Topologically consistent source domain graph ,in Will , Embed the target node attribute structure to ensure isomorphism.

[0043] S33, Around the fiber principal vector With the fixture coordinate system ,right Perform small angle rotation Both geometric enhancement and mirror reflection are performed, and the edge connection vectors are updated simultaneously. With border rights Generate additional source map All enhancement results will be merged into the source domain graph pool. This enhances the diversity of source domain representations.

[0044] S34, Construct a shared encoder First, the source pool With target map Node feature matrix Perform joint training to obtain node embeddings :

[0045]

[0046] in, For nodes The latent space embedding vector; The node feature vector defined in step S2; , This is the encoder weight matrix; It is a linear rectification activation function with leakage parameters.

[0047] After completing the above training, freeze. Preserve with fixed geometrically related projections The learnable channel provides a stable geometric base for adversarial alignment.

[0048] S35, at encoder output Graph-level convergence vector Access Domain Identification Subnet The subnet outputs domain labels in a binary classification format. By minimizing the adversarial loss source domain, the obfuscation term in the project's target domain is identified.

[0049]

[0050] in, To counteract domain alignment loss; , For graph-level embedding of the source and target domains; For domain discrimination subnet parameter set .

[0051] by The alternating updates of the two-player game drive the encoder to erase domain differences, while the discriminant subnet continuously distinguishes domain labels, achieving convergence of high-order feature distributions.

[0052] S36. To suppress geometric dominance drift during the adversarial training phase, at each mapping node... Calculate statistical difference ,in , These are the batch averages for temperature and curvature, respectively; Weighted Inclusion This enables node-level statistical matching, prompting the temperature difference and morphological benchmark of the two domains to converge simultaneously.

[0053] S37, When the accuracy of local discrimination drops to and At this point, adversarial training is stopped, all encoder weights are frozen, the source domain equalization parameters are mapped to the target domain feature map, and an alignment map with unified semantics is output. This is used for subsequent generation of adversarial graph transfer networks.

[0054] Optionally, S4 includes the following steps:

[0055] S41, in the alignment diagram The generator and discriminator are directly connected; the first layer of the generator constructs a fiber orientation adaptive graph convolution, which differs from a general adaptive graph convolution in that it is based on the node fiber vector. Dynamically modulate the original edge weights Obtain the right to amend :

[0056]

[0057] in, The basic edge weights are calculated in step S2; This is the fiber gain coefficient; For nodes Normalized fiber orientation vector; Connect the nodes with unit vectors .

[0058] According to the above formula, the convolution kernel amplifies the receptive field in the direction parallel to the fiber and suppresses redundant diffusion in the vertical direction; the final residual Laplacian enhancement layer adds the multi-scale Laplacian features to the backbone output element by element, enhancing the warping details in the extremely small radius region.

[0059] S42. Training phase discriminator parameters locking Only high-fidelity finite element warp field The generator is trained through ten rounds of reconstruction; the loss term is a mixture of node mean squared error and local gradient difference to ensure the generator output... Quickly capture the overall warping trend and avoid gradient oscillations caused by early counter-offsets.

[0060] S43. After preheating, the thawing detector sends... Synchronous input and The discriminator calculates the peak stress difference at the nodes, the second-order Laplace smoothness of the deformation field, and the strain energy difference across the entire image, and outputs a domain confidence score. The above scores will be fed back to the generator as dynamic weights, driving the alignment of feature details.

[0061] S44. During the alternating update process, the overall objective function is set as follows:

[0062]

[0063] in, This refers to the error in the reconstruction of the nodal field. To combat cross-entropy loss, drive confuse ; This represents the sum of squares of the peak stress differences; The residual is a second-order Laplace. This refers to the global strain energy closure error. For stage-adaptive weights, based on Dynamic adjustment.

[0064] Through backpropagation of the above equation, when When outputting high confidence for three consecutive rounds, the residual Laplace channel weights are automatically increased to promote rapid convergence of details.

[0065] S45. Real-time monitoring of the fiber parallel error curve; when the error descent gradient in this direction falls below a set threshold, immediately relax the weight decay and increase the learning rate. This allows the adaptive convolution kernel to further enhance its sensitivity to fiber principal direction information, ensuring that the anisotropic response is fully learned.

[0066] S46. After each iteration, calculate the energy closure error on the independent verification graph. ;like Exceeding the project limit That is, back up the generator to the previous stable weight, and... reduce This is to prevent opposing terms from weakening the physical conservation.

[0067] S47, When the comprehensive error The decline was less than 5 times in five consecutive rounds. and At this point, freeze all parameters of the generator and discriminator, and export the trained generator. The generator provides a reliable initial deformation field for the multi-objective joint optimization in step S5, while ensuring node-level physical consistency.

[0068] Optionally, S5 includes the following steps:

[0069] S51, Reader / Discriminator The latest inference results yielded the node smoothness residual vectors respectively. Peak stress difference vector scalar of overall energy deviation The three types of physical indicators are concatenated into a physical consistency vector:

[0070]

[0071] This provides a quantitative benchmark for subsequent weight scheduling.

[0072] in, The node's second-order Laplace smooth residual vector; The peak stress difference vector at the nodes; This represents the overall strain energy deviation. It is the vector norm 2; It is the infinite norm of the vector.

[0073] S52, Output warp field in generator With high-fidelity samples Calculate node-level reconstruction error and the physical consistency vector of step S51 and counter-matching error Fusion, constructing a multi-objective loss function:

[0074]

[0075] in, The total loss; For adaptive weights; The total number of nodes; For the generator Nodal equivalent stress; To design the allowable stress threshold; This is the energy constraint scaling factor; The generator outputs an equivalent variable energy; This represents the theoretical minimum strain energy calculated according to engineering standards.

[0076] S53, using a length of Sliding window record Historical mean and variance; batch normalization is used to scale the magnitude of each loss gradient to... This prevents a single component from dominating backpropagation and establishes a unified numerical scale for weighted scheduling.

[0077] S54. Design a dynamic weight scheduler based on the physical consistency vector. Real-time changes and adjustments Specifically, an exponential ratio update is used:

[0078]

[0079] in, For the first Round Weight; This is the scheduling gain coefficient; For the first in the sliding window The latest mean squared error of the components; This is the total mean square error of the three components, used for normalization.

[0080] When physical indicators exceed the threshold Increase, make This results in exponential amplification; if the visual reconstruction error remains high, then... The weights are increased accordingly, thereby dynamically balancing visual accuracy and physical plausibility during training.

[0081] S55. Perform alternating optimization, freezing the discriminator in each outer iteration cycle and only using... Reverse update generator; if detected If the oscillation exceeds the set threshold, the reconstruction channel gradient will be temporarily frozen for two rounds to prevent the fiber principal direction details from being weakened due to excessive smoothing.

[0082] S56, every other time The wheel is evaluated on an independent verification plot to assess the maximum nodal stress. Energy difference with the whole image When both are below the engineering tolerance line Then, lock the current weight range and lower the learning rate. We have now entered the detailed refinement stage, focusing on optimizing local warping peaks and microscale stress concentrations.

[0083] S57, when continuous All rounds of verification are satisfied , and The rate of change is lower than At that point, training is terminated, and the generator's final weights are saved. With the dual guarantees of morphological consistency and physical reliability, the weights directly serve as a highly reliable initial field for subsequent local fine-tuning and process path optimization.

[0084] Optionally, S6 includes the following steps:

[0085] S61. Obtain the geometric-physical feature map of the structure to be predicted. Input freeze generator Obtained the first edition of the warp field Simultaneously calculate the nodal gradients. Equivalent stress estimation with anisotropy And cache it as a risk assessment benchmark.

[0086] S62, Synthesize curvature jump variables for each node Predicted peak stress ratio Historical error replay item Construct a risk index:

[0087]

[0088] in For normalized weights. Sort by highest to lowest, and extract the first few. Nodes are aggregated into a high-risk subdomain set based on topological connectivity. .

[0089] S63. Process each subdomain sequentially. The local adaptive mesher is invoked to generate a refined mesh based on the nodal curvature gradient, and then a second-level low-fidelity heat transfer-elasticity solver is driven, with the same furnace temperature curve as input. Output local warp correction amount The correction results are mapped to the original image node indices using nearest neighbor interpolation.

[0090] S64. Enable a writable channel in the graph network refinement layer. After writing back to the corresponding node, use a Gaussian kernel. The diffusion extends two layers of neighborhood along the graph edge, using a weight decay factor during the diffusion process. Ensure a smooth transition between local and global environments.

[0091] S65. Keep the backbone parameters frozen, only unlock the refinement layer weights. Select the difference before and after correction. As a monitoring signal, based on mini-batch incremental gradient pairs conduct The network is updated step by step to add local features to its memory, thus avoiding global overfitting.

[0092] S66. Obtain the second version of the warp field using updated network inference. If the residual error of all high-risk nodes Less than the threshold If the threshold scaling factor is correct, the final warp field will be output; otherwise, the threshold scaling factor will be adjusted. Multiply Then return to S62 and continue iterating until the accuracy requirements are met.

[0093] S67. The image that has been fine-tuned. Local correction records With updated weights Unified archiving to the source domain sample pool; triggering main network retraining during daily offline phases, using the forgetting rate. The experience replay strategy continuously injects new local knowledge to achieve the continuous evolution of overall prediction accuracy.

[0094] The beneficial effects of this invention are:

[0095] This invention introduces three types of entities—CAD geometric nodes, material parameter nodes, and thermal-mechanical state nodes—into a temporal heterogeneous graph, along with three types of semantic edges: functional block partitioning, principal heat flow tracing, and initial warp field. Combined with a time-temperature dual index, it constructs a complete coding framework for the evolution from heat conduction to solidification shrinkage and then to warping. This structure enables fine-grained modeling of warping evolution patterns under different time periods, different functional blocks, and multiple semantic relationships, effectively overcoming the limitations of existing methods in extracting heterogeneous thermal-mechanical coupling features and capturing timeliness.

[0096] This invention designs a domain alignment strategy based on generative adversarial graph transfer networks and node-level physical prior statistical matching: First, a high-fidelity source domain graph that is similar to the target component in geometry, plying, and process is retrieved and mapped into an isomorphic graph. Then, within a shared encoder-domain discriminant subnet framework, adversarial and statistical matching of node temperature difference and curvature distribution is used to map historical simulated physical fields into dynamic domain prior labels and align them with the features of the target graph. This method significantly improves the consistency and generalization ability of cross-component warp field transfer and enhances the model's real-time perception of physical constraints.

[0097] This invention constructs a fiber orientation adaptive graph convolution generator and a multi-task physics-sensitive discriminator, combining contextual information such as temperature, curvature, and stress fields for adversarial training and multi-objective joint optimization. At the generator level, edge weights are dynamically modulated to enhance the fiber principal direction response; at the discriminator level, nodal peak stress difference, second-order Laplace residual, and full-graph strain energy closure error are introduced as discriminative priors, and dynamic weight scheduling ensures energy conservation. Combined with a risk subdomain local adaptive mesh refinement and incremental fine-tuning mechanism, this invention not only achieves accurate transfer mapping of high-fidelity warped fields but also effectively solves the problems of poor sensitivity to physical details and insufficient ability to correct local extreme warping points in previous adversarial transfer models. Attached Figure Description

[0098] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0099] Figure 1 This is a flowchart of a warping prediction and optimization method for complex carbon fiber composite suspended structures proposed in this invention.

[0100] Figure 2 This is a flowchart of the connection generator and discriminator for a warping prediction and optimization method for complex carbon fiber composite suspended structures proposed in this invention.

[0101] Figure 3 This diagram illustrates a comparison between the proposed warpage prediction and optimization method for complex carbon fiber composite suspended structures and existing technologies.

[0102] Figure 4 This is a schematic diagram of the final deformation of a warping prediction and optimization method for complex carbon fiber composite suspended structures proposed in this invention. Detailed Implementation

[0103] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0104] refer to Figure 1 A method for predicting and optimizing warpage in complex carbon fiber composite cantilever structures includes the following steps:

[0105] S1. Read the CAD geometry file and fiber layup description file of the target component, and perform heat flow main path tracing along the fiber principal direction. When the tracing path encounters a dense area of ​​support points or a transition position of the first derivative of the curvature, the tracing is terminated. The component is divided into multiple functional blocks that maintain the independence of anisotropic heat conduction according to the terminated path. For each functional block, the fixture contact line is detected. If the same fixture contact line crosses multiple blocks, the Boolean union of the crossing blocks is performed or the fixture contact line is re-trimmed until the fixture contact line is completely located within a single block. Query the material comparison matrix to find the fiber layup angle sequence and interlayer thickness. The temperature sequence and resin type are simplified into in-plane equivalent thermal expansion coefficient, equivalent elastic modulus, equivalent density, and specific heat capacity, and packaged into a simplified material card; a three-dimensional adaptive mesh is generated based on the estimated temperature gradient, and if the total number of nodes exceeds the set upper limit, the mesh size is increased synchronously; the furnace temperature curve and the simplified material card are input into the steady-state heat transfer fast solver to obtain the node temperature, and the node thermal strain is calculated based on the node temperature difference and the curing shrinkage ratio. Then, the initial value field of block warping is obtained by integrating along the thickness direction. The initial value field of warping is subjected to weighted Laplace smoothing and normalization to form a priori package containing node coordinates, temperature field, and warping field;

[0106] S2. Read the prior package, first rearrange the node coordinates according to the functional block number, then according to the order of the X component and Y component; at each node, use quadratic surface least squares fitting based on a ring neighborhood to obtain the three-dimensional principal curvature, and use the Voronoi area kernel method to obtain the node control area; project the node normal vector to the fiber principal vector to obtain the node fiber layup angle, and combine the principal curvature, control area, fiber layup angle, local laminate thickness, peak temperature, cooling stage temperature slope and warpage initial value into a node feature vector; within the same block, establish connections between node pairs whose Euclidean distance does not exceed the cutoff radius and calculate the edge weights based on the node heat flow principal vector and temperature difference; between cross-block node pairs, establish connections only when the temperature difference or warpage gradient exceeds the threshold and use the same edge weight calculation method; perform zero-mean unit variance normalization on the node feature vector to generate a geometric-physical feature map containing a node set, a weighted edge set, a node feature matrix and a sparse edge weight matrix.

[0107] S3. Retrieve samples from historical high-fidelity thermo-elastic coupling simulation archives whose Euclidean distance to the feature vector of the target component does not exceed a preset threshold to construct a source domain pool. Map each simulation record in the source domain pool to a source domain graph isomorphic to the target graph according to the node index rearrangement table and edge weight rules in step S2. Perform small-angle rotation and mirror reflection enhancement on the source domain graph around the fiber principal direction and the fixture coordinate system and then merge it into the source domain graph pool. Construct a shared encoder to jointly train the node feature matrix of the source domain graph pool and the target graph to obtain the node latent space embedding. Connect it to the domain discrimination subnet. Eliminate the difference between the source domain and the target domain through adversarial loss and apply temperature difference and curvature statistical matching at the node level. When the domain discrimination accuracy drops to a set threshold and the node statistical difference converges, output the alignment map.

[0108] S4. Connect the generator and discriminator on the alignment graph. The first layer of the generator uses fiber orientation adaptive graph convolution to dynamically modulate the edge weights based on the node fiber vectors. The last layer is set with a residual Laplace enhancement layer. First, lock the discriminator parameters and preheat the generator using a high-fidelity warp field. Then, unfreeze the discriminator and alternately update the generator and discriminator. The discriminator outputs the node peak stress difference, the second-order Laplace smoothness of the deformation field, and the global strain energy difference as dynamic weights to be fed back to the generator. Train the generator with a comprehensive objective function that includes node reconstruction error, adversarial cross-entropy, sum of squares of peak stress difference, second-order Laplace residual, and global strain energy closure error until the comprehensive error decreases by less than a preset threshold for five consecutive rounds and the energy closure error meets the engineering requirements. Then, export the trained generator.

[0109] S5. Based on the latest inference results of the discriminator, obtain the node second-order Laplace smooth residual vector, the node peak stress difference vector, and the full-image strain energy deviation scalar to construct a physical consistency vector. This vector is then fused with the node-level reconstruction error and adversarial matching error of the warped field output by the generator to form a multi-objective loss function containing reconstruction terms, adversarial terms, peak stress constraint terms, and energy constraint terms. A sliding window is used to record historical mean and variance, and the magnitude of each gradient is normalized. The weights of each sub-loss are dynamically adjusted through an exponential proportional update strategy. Oscillations are monitored during alternating optimization, and the gradient of the reconstruction channel is frozen when necessary. When the maximum node stress and the full-image strain energy difference in the validation set are both below the engineering tolerance line and the rate of change converges, all parameters of the generator and discriminator are frozen.

[0110] S6. Input the geometric-physical feature map of the structure to be predicted into the frozen generator to obtain the first version of the warp field. Calculate the node gradient and anisotropic equivalent stress estimation, and calculate the risk index based on the node curvature jump variable, the peak predicted stress ratio, and the historical error replay term. Cluster high-risk nodes according to topological connectivity to form several subdomains. Perform local adaptive mesh refinement on each high-risk subdomain and call the low-fidelity heat transfer-elasticity solver to output the local warp correction amount. Use nearest neighbor interpolation to write back the original image nodes and spread two layers of neighborhood along the edge on the image. Unlock the refinement layer weights and use the difference before and after correction as the supervision signal for incremental learning and inference again. If the residual error of all high-risk nodes is lower than the threshold, output the final warp field. Otherwise, scale the threshold and repeat the above high-risk subdomain correction and incremental learning until the accuracy requirement is met. Finally, archive the fine-tuned feature map, local correction record, and updated weights to the source domain sample pool for offline retraining.

[0111] S1 includes the following sub-steps:

[0112] S11. Read the CAD geometry file and the corresponding layup description file, and perform main heat flow path tracing according to the fiber principal direction and the continuous direction of the outer curvature; when the tracing path encounters a dense area of ​​support points or a transition position of the first derivative of the outer curvature, the tracing is terminated immediately, thereby obtaining a preliminary set of functional blocks that maintain the independence of anisotropic heat conduction. .

[0113] S12. For each preliminary functional block, perform a spatial overlap determination between its boundary surface and the fixture contact line. If the same fixture contact line is found to span multiple blocks, perform geometric Boolean operations on the spanning blocks to parallelize them into a new single block. Alternatively, the clamp contact lines can be re-cut until all clamp contact lines are completely within a single block, ensuring that the boundary conditions for subsequent low-fidelity solutions are closed.

[0114] S13, Block Read fiber layup angle sequence Interlayer thickness sequence and resin type By querying the preset material comparison matrix, the in-plane equivalent thermal expansion coefficient of the block can be obtained. Equivalent elastic modulus equivalent density and specific heat capacity And package it into a simplified material card This reduces the dimensionality of subsequent solution parameters.

[0115] S14, in the block Internally based on the predicted temperature gradient Generate a 3D adaptive mesh. Mesh nodes. Size Calculated using the following formula:

[0116]

[0117] in, As a global baseline length scale; This is the gradient adjustment coefficient; For nodes Temperature gradient prediction model at the location; For nodes The grid size at that location.

[0118] Statistics of all The corresponding total number of nodes, and the set upper limit. Compare; if the total exceeds Then according to The incremental method increases synchronously. The value is maintained until the number of nodes meets the upper limit requirement.

[0119] S15, Block Input furnace temperature curve and materials card The nodal temperatures were obtained by solving the problem. Combined with curing shrinkage ratio With node temperature difference Calculate nodal thermal strain .in This is a reference temperature.

[0120] S16. Apply thermal strain to the node along the thickness direction. Integrate to obtain the spontaneous warping of the block. :

[0121]

[0122] For nodes The initial value of the warping displacement; For blocks The equivalent bending stiffness factor; For blocks The total thickness; For thickness position node Thermal strain.

[0123] The integral results are mapped back to the coarse-grained grid nodes through linear interpolation to form the initial value field for block warping. .

[0124] S17. Perform weighted Laplace smoothing and normalization on the warped initial values ​​of all blocks to remove discontinuities in cross-block values; then, set the node coordinates... Temperature field With normalized warp field Encapsulated as a priori package .

[0125] S2 includes the following sub-steps:

[0126] S21. Read the node coordinate set given in step S1. Temperature field With warped field Sort by functional block number first, then within each functional block, use coordinates... of - Component order, - The component order is rearranged in two stages to ensure that neighboring nodes remain continuous in the index space, eliminating the geometric coherence disruption caused by random grid generation.

[0127] S22, at each node At this point, the triaxial principal curvature is calculated using a quadratic surface least squares fitting method based on the set of its one-ring neighborhood nodes. , , Simultaneously, the nodal control area is obtained using the Voronoi area subdivision method. The curvature-area pair jointly describes the degree of abrupt changes in local stiffness, providing a geometric basis for determining warp sensitivity.

[0128] S23, Adjust node normals Projected onto the block fiber principal vector The fiber layup angle at the node is obtained from the included projection angle. Combined with the total interlayer thickness ,Will , and Write node properties to explicitly encode the anisotropic stress response.

[0129] S24. Synchronously load the temperature rise curve of the corresponding node. Record peak temperature and the slope of linear regression during the cooling phase Following the aforementioned geometric and material properties, the complete feature vector of a node is defined as follows:

[0130]

[0131] in, , , For nodes exist , , Principal curvature in the direction; Control area for nodes based on Voronoi area allocation; For the fiber layup angle at the node; The thickness of the local laminate at the node; This refers to the peak temperature at the node. The temperature slope during the node cooling phase; The initial value of node warping is calculated in step S1.

[0132] S25. Within the same functional block, for each node Select a value whose Euclidean distance does not exceed the cutoff radius. All neighboring nodes Forming a connecting edge;

[0133] Combining the principal direction vector of nodal heat flow , With temperature difference Calculate the local edge weights:

[0134]

[0135] in, , The coordinates of the node; The distance attenuation scale; This is the gain coefficient in the heat flow direction; Connect the nodes with unit vectors; This is the temperature difference gain coefficient; , This represents the principal direction vector of heat flow at the node; The temperature difference between nodes; The radius of the cutoff point is the radius of the connecting edge.

[0136] S26. In cross-functional block node pairs First, check the temperature difference. With warp gradient ;when or Cross-block connections are established only when the weights are calculated based on Equation S25, in order to highlight potential warping jump points.

[0137] S27. For all node feature vectors Apply zero-mean unit variance normalization; for all edge weights Layered caching by functional blocks allows subsequent networks to distinguish between local and cross-block coupling strength. (Geometric-physical feature map) Construction complete, including For a set of nodes, For weighted edge set, For node feature matrix, It is a sparse edge weight matrix.

[0138] S3 includes the following sub-steps:

[0139] S31. Retrieve high-fidelity thermo-elastic coupling simulation files of completed carbon fiber suspension components from the enterprise database, and search for geometric reference dimensions respectively. Layer sequence With furnace temperature curve Calculate the feature vector corresponding to the target part. Euclidean distance; only retain distances not exceeding a threshold. The samples are used to construct the initial pool of the source domain. Reduce irrelevant structural noise interference during alignment training.

[0140] S32, For the initial pool of the source domain For each simulation record, extract the node temperature. With nodal strain Based on the node index rearrangement table and edge weight rules in step S2, a point-by-point mapping is generated to... Topologically consistent source domain graph ,in Will , Embed the target node attribute structure to ensure isomorphism.

[0141] S33, Around the fiber principal vector With the fixture coordinate system ,right Perform small angle rotation Both geometric enhancement and mirror reflection are performed, and the edge connection vectors are updated simultaneously. With border rights Generate additional source map All enhancement results will be merged into the source domain graph pool. This enhances the diversity of source domain representations.

[0142] S34, Construct a shared encoder First, the source pool With target map Node feature matrix Perform joint training to obtain node embeddings :

[0143]

[0144] in, For nodes The latent space embedding vector; The node feature vector defined in step S2; , This is the encoder weight matrix; It is a linear rectification activation function with leakage parameters.

[0145] After completing the above training, freeze. Preserve with fixed geometrically related projections The learnable channel provides a stable geometric base for adversarial alignment.

[0146] S35, at encoder output Graph-level convergence vector Access Domain Identification Subnet The subnet outputs domain labels in a binary classification format. By minimizing the adversarial loss source domain, the obfuscation term in the project's target domain is identified.

[0147]

[0148] in, To counteract domain alignment loss; , For graph-level embedding of the source and target domains; For domain discrimination subnet parameter set .

[0149] by The alternating updates of the two-player game drive the encoder to erase domain differences, while the discriminant subnet continuously distinguishes domain labels, achieving convergence of high-order feature distributions.

[0150] S36. To suppress geometric dominance drift during the adversarial training phase, at each mapping node... Calculate statistical difference ,in , These are the batch averages for temperature and curvature, respectively; Weighted Inclusion This enables node-level statistical matching, prompting the temperature difference and morphological benchmark of the two domains to converge simultaneously.

[0151] S37, When the accuracy of local discrimination drops to and At this point, adversarial training is stopped, all encoder weights are frozen, the source domain equalization parameters are mapped to the target domain feature map, and an alignment map with unified semantics is output. This is used for subsequent generation of adversarial graph transfer networks.

[0152] S4 includes the following sub-steps:

[0153] S41, in the alignment diagram The generator and discriminator are directly connected; the first layer of the generator constructs a fiber orientation adaptive graph convolution, which differs from a general adaptive graph convolution in that it is based on the node fiber vector. Dynamically modulate the original edge weights Obtain the right to amend :

[0154]

[0155] in, The basic edge weights are calculated in step S2; This is the fiber gain coefficient; For nodes Normalized fiber orientation vector; Connect the nodes with unit vectors .

[0156] According to the above formula, the convolution kernel amplifies the receptive field in the direction parallel to the fiber and suppresses redundant diffusion in the vertical direction; the final residual Laplacian enhancement layer adds the multi-scale Laplacian features to the backbone output element by element, enhancing the warping details in the extremely small radius region.

[0157] S42. Training phase discriminator parameters locking Only high-fidelity finite element warp field The generator is trained through ten rounds of reconstruction; the loss term is a mixture of node mean squared error and local gradient difference to ensure the generator output... Quickly capture the overall warping trend and avoid gradient oscillations caused by early counter-offsets.

[0158] S43. After preheating, the thawing detector sends... Synchronous input and The discriminator calculates the peak stress difference at the nodes, the second-order Laplace smoothness of the deformation field, and the strain energy difference across the entire image, and outputs a domain confidence score. The above scores will be fed back to the generator as dynamic weights, driving the alignment of feature details.

[0159] S44. During the alternating update process, the overall objective function is set as follows:

[0160]

[0161] in, This refers to the error in the reconstruction of the nodal field. To combat cross-entropy loss, drive confuse ; This represents the sum of squares of the peak stress differences; The residual is a second-order Laplace. This refers to the global strain energy closure error. For stage-adaptive weights, based on Dynamic adjustment.

[0162] Through backpropagation of the above equation, when When outputting high confidence for three consecutive rounds, the residual Laplace channel weights are automatically increased to promote rapid convergence of details.

[0163] S45. Real-time monitoring of the fiber parallel error curve; when the error descent gradient in this direction falls below a set threshold, immediately relax the weight decay and increase the learning rate. This allows the adaptive convolution kernel to further enhance its sensitivity to fiber principal direction information, ensuring that the anisotropic response is fully learned.

[0164] S46. After each iteration, calculate the energy closure error on the independent verification graph. ;like Exceeding the project limit That is, back up the generator to the previous stable weight, and... reduce This is to prevent opposing terms from weakening the physical conservation.

[0165] S47, When the comprehensive error The decline was less than 5 times in five consecutive rounds. and At this point, freeze all parameters of the generator and discriminator, and export the trained generator. The generator provides a reliable initial deformation field for the multi-objective joint optimization in step S5, while ensuring node-level physical consistency.

[0166] S5 includes the following sub-steps:

[0167] S51, Reader / Discriminator The latest inference results yielded the node smoothness residual vectors respectively. Peak stress difference vector scalar of overall energy deviation The three types of physical indicators are concatenated into a physical consistency vector:

[0168]

[0169] This provides a quantitative benchmark for subsequent weight scheduling.

[0170] in, The node's second-order Laplace smooth residual vector; The peak stress difference vector at the nodes; This represents the overall strain energy deviation. It is the vector norm 2; It is the infinite norm of the vector.

[0171] S52, Output warp field in generator With high-fidelity samples Calculate node-level reconstruction error and the physical consistency vector of step S51 and counter-matching error Fusion, constructing a multi-objective loss function:

[0172]

[0173] in, The total loss; For adaptive weights; The total number of nodes; For the generator Nodal equivalent stress; To design the allowable stress threshold; This is the energy constraint scaling factor; The generator outputs an equivalent variable energy; This represents the theoretical minimum strain energy calculated according to engineering standards.

[0174] S53, using a length of Sliding window record Historical mean and variance; batch normalization is used to scale the magnitude of each loss gradient to... This prevents a single component from dominating backpropagation and establishes a unified numerical scale for weighted scheduling.

[0175] S54. Design a dynamic weight scheduler based on the physical consistency vector. Real-time changes and adjustments Specifically, an exponential ratio update is used:

[0176]

[0177] in, For the first Round Weight; This is the scheduling gain coefficient; For the first in the sliding window The latest mean squared error of the components; This is the total mean square error of the three components, used for normalization.

[0178] When physical indicators exceed the threshold Increase, make This results in exponential amplification; if the visual reconstruction error remains high, then... The weights are increased accordingly, thereby dynamically balancing visual accuracy and physical plausibility during training.

[0179] S55. Perform alternating optimization, freezing the discriminator in each outer iteration cycle and only using... Reverse update generator; if detected If the oscillation exceeds the set threshold, the reconstruction channel gradient will be temporarily frozen for two rounds to prevent the fiber principal direction details from being weakened due to excessive smoothing.

[0180] S56, every other time The wheel is evaluated on an independent verification plot to assess the maximum nodal stress. Energy difference with the whole image When both are below the engineering tolerance line Then, lock the current weight range and lower the learning rate. We have now entered the detailed refinement stage, focusing on optimizing local warping peaks and microscale stress concentrations.

[0181] S57, when continuous All rounds of verification are satisfied , and The rate of change is lower than At that point, training is terminated, and the generator's final weights are saved. With the dual guarantees of morphological consistency and physical reliability, the weights directly serve as a highly reliable initial field for subsequent local fine-tuning and process path optimization.

[0182] S6 includes the following sub-steps:

[0183] S61. Obtain the geometric-physical feature map of the structure to be predicted. Input freeze generator Obtained the first edition of the warp field Simultaneously calculate the nodal gradients. Equivalent stress estimation with anisotropy And cache it as a risk assessment benchmark.

[0184] S62, Synthesize curvature jump variables for each node Predicted peak stress ratio Historical error replay item Construct a risk index:

[0185]

[0186] in For normalized weights. Sort by highest to lowest, and extract the first few. Nodes are aggregated into a high-risk subdomain set based on topological connectivity. .

[0187] S63. Process each subdomain sequentially. The local adaptive mesher is invoked to generate a refined mesh based on the nodal curvature gradient, and then a second-level low-fidelity heat transfer-elasticity solver is driven, with the same furnace temperature curve as input. Output local warp correction amount The correction results are mapped to the original image node indices using nearest neighbor interpolation.

[0188] S64. Enable a writable channel in the graph network refinement layer. After writing back to the corresponding node, use a Gaussian kernel. The diffusion extends two layers of neighborhood along the graph edge, using a weight decay factor during the diffusion process. Ensure a smooth transition between local and global environments.

[0189] S65. Keep the backbone parameters frozen, only unlock the refinement layer weights. Select the difference before and after correction. As a monitoring signal, based on mini-batch incremental gradient pairs conduct The network is updated step by step to add local features to its memory, thus avoiding global overfitting.

[0190] S66. Obtain the second version of the warp field using updated network inference. If the residual error of all high-risk nodes Less than the threshold If the threshold scaling factor is correct, the final warp field will be output; otherwise, the threshold scaling factor will be adjusted. Multiply Then return to S62 and continue iterating until the accuracy requirements are met.

[0191] S67. The image that has been fine-tuned. Local correction records With updated weights Unified archiving to the source domain sample pool; triggering main network retraining during daily offline phases, using the forgetting rate. The experience replay strategy continuously injects new local knowledge to achieve the continuous evolution of overall prediction accuracy.

[0192] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting and optimizing warpage in complex carbon fiber composite cantilever structures, characterized in that, Includes the following steps: S1: Read in CAD and layup, trace the main heat flow along the fiber to divide anisotropic blocks, generate equivalent material cards, perform adaptive mesh steady-state solution and obtain warpage prior field; S2: Rearrange nodes, locally fit principal curvature and control area, extract fiber angle, layer thickness, temperature, and warpage features, and construct a weighted adjacency graph; S3: Retrieve historical high-fidelity samples, map and enhance them into source domain graphs, and conduct adversarial training between the joint encoder and the domain discriminator to achieve node-level alignment; S4: Based on the alignment map, a generator and discriminator are constructed, and after preheating, alternating adversarial optimization is performed, with dynamic weighted stress and smoothing of errors until convergence; S5: Construct a multi-objective loss mechanism including reconstruction, adversarial, peak stress, and energy constraints; dynamically normalize the gradient and iterate to tolerance by adjusting the weights. S6: Freeze the initial warping of the generator inference, evaluate node risk clustering of high-risk subdomains, refine the local grid and iterate incrementally until the model is tolerated and archived.

2. The method for predicting and optimizing warpage of complex carbon fiber composite cantilever structures according to claim 1, characterized in that, S1 includes the following steps: S11. Read the CAD geometry file and the corresponding layup description file, and perform main heat flow path tracing according to the fiber principal direction and the continuous direction of the outer curvature; when the tracing path encounters a dense area of ​​support points or a transition position of the first derivative of the outer curvature, the tracing is terminated immediately, thereby obtaining a preliminary set of functional blocks that maintain the independence of anisotropic heat conduction. ; S12. For each preliminary functional block, perform a spatial overlap determination between its boundary surface and the fixture contact line. If the same fixture contact line is found to span multiple blocks, perform geometric Boolean operations on the spanning blocks to parallelize them into a new single block. Alternatively, the clamp contact lines can be re-cut until all clamp contact lines are completely within a single block, ensuring that the boundary conditions for subsequent low-fidelity solutions are closed. S13, Block Read fiber layup angle sequence Interlayer thickness sequence and resin type By querying the preset material comparison matrix, the in-plane equivalent thermal expansion coefficient of the block can be obtained. Equivalent elastic modulus equivalent density and specific heat capacity And package it into a simplified material card This reduces the dimensionality of subsequent solution parameters; S14, in the block Internally based on the predicted temperature gradient Generate a 3D adaptive mesh, mesh nodes Size Calculated using the following formula: ； in, As a global baseline length scale; This is the gradient adjustment coefficient; For nodes Temperature gradient prediction model at the location; For nodes Grid size at the location; Statistics of all The corresponding total number of nodes, and the set upper limit. Compare; if the total exceeds Then according to The incremental method increases synchronously. The value is maintained until the number of nodes meets the upper limit requirement; S15, Block Input furnace temperature curve and materials card The nodal temperatures were obtained by solving the problem. Combined with curing shrinkage ratio With node temperature difference Calculate nodal thermal strain ,in For reference temperature; S16. Apply thermal strain to the node along the thickness direction. Integrate to obtain the spontaneous warping of the block. : ； For nodes The initial value of the warping displacement; For blocks The equivalent bending stiffness factor; For blocks The total thickness; For thickness position node Thermal strain; The integral results are mapped back to the coarse-grained grid nodes through linear interpolation to form the initial value field for block warping. ; S17. Perform weighted Laplace smoothing and normalization on the warped initial values ​​of all blocks to remove discontinuities in cross-block values; then, set the node coordinates... Temperature field With normalized warp field Encapsulated as a priori package .

3. The method for predicting and optimizing warpage of complex carbon fiber composite suspended structures according to claim 1, characterized in that, S2 includes the following steps: S21. Read the node coordinate set given in step S1. Temperature field With warped field Sort by functional block number first, then within each functional block, use coordinates... of - Component order, - The component order is rearranged in two stages to ensure that neighboring nodes remain continuous in the index space, eliminating the geometric coherence disruption caused by random grid generation; S22, at each node At this point, the triaxial principal curvature is calculated using a quadratic surface least squares fitting method based on the set of its one-ring neighborhood nodes. , , Simultaneously, the nodal control area is obtained using the Voronoi area subdivision method. The curvature-area pair jointly describes the degree of abrupt change in local stiffness, providing a geometric basis for determining warp sensitivity; S23, adjust the node normals Projected onto the block fiber principal vector The fiber layup angle at the node is obtained from the included projection angle. Combined with the total interlayer thickness ,Will , and Write node properties to explicitly encode the anisotropic stress response; S24. Synchronously load the temperature rise curve of the corresponding node. Record peak temperature and the slope of linear regression during the cooling phase Following the aforementioned geometric and material properties, the complete feature vector of a node is defined as follows: ； in, , , For nodes exist , , Principal curvature in the direction; Control area for nodes based on Voronoi area allocation; For the fiber layup angle at the node; The thickness of the local laminate at the node; This refers to the peak temperature at the node. The temperature slope during the node cooling phase; The initial value of node warping calculated in step S1; S25. Within the same functional block, for each node Select a value whose Euclidean distance does not exceed the cutoff radius. All neighboring nodes Form a connecting edge; Combining the principal direction vector of nodal heat flow , With temperature difference Calculate the local edge weights: ； in, , The coordinates of the node; The distance attenuation scale; This is the gain coefficient in the heat flow direction; Connect the nodes with unit vectors; This is the temperature difference gain coefficient; , This represents the principal direction vector of heat flow at the node; The temperature difference between nodes; The radius of the cutoff point for the connected edge; S26. In cross-functional block node pairs First, check the temperature difference. With warp gradient ;when or Cross-block connections are established only when the weights are calculated based on Equation S25, in order to highlight potential warping jump points; S27. For all node feature vectors Apply zero-mean unit variance normalization; for all edge weights Layered caching by functional blocks allows subsequent networks to distinguish between local and cross-block coupling strength, and geometric-physical feature maps. Construction complete, including For a set of nodes, For weighted edge set, For node feature matrix, It is a sparse edge weight matrix.

4. The method for predicting and optimizing warpage of complex carbon fiber composite cantilever structures according to claim 1, characterized in that, S3 includes the following steps: S31. Retrieve high-fidelity thermo-elastic coupling simulation files of completed carbon fiber suspension components from the enterprise database, and search for geometric reference dimensions respectively. Layer sequence With furnace temperature curve Calculate the feature vector corresponding to the target part. Euclidean distance; only retain distances not exceeding a threshold. Using samples, construct the initial pool of the source domain. Reduce irrelevant structural noise interference during training alignment; S32, For the initial pool of the source domain For each simulation record, extract the node temperature. With nodal strain Based on the node index rearrangement table and edge weight rules in step S2, a point-by-point mapping is generated to... Topologically consistent source domain graph ,in Will , Embed the target node's attribute structure to ensure isomorphism; S33, Around the fiber principal vector With the fixture coordinate system ,right Perform small angle rotation Both geometric enhancement and mirror reflection are performed, and the edge connection vectors are updated simultaneously. With border rights Generate additional source map All enhancement results will be merged into the source domain graph pool. Enhance the diversity of source domain representations; S34, Construct a shared encoder First, the source pool With target map Node feature matrix Perform joint training to obtain node embeddings : ； in, For nodes The latent space embedding vector; The node feature vector defined in step S2; , This is the encoder weight matrix; It is a linear rectified activation function with leakage parameters; After completing the above training, freeze. Preserve the fixed geometrically related projection. The learnable channel provides a stable geometric base for adversarial alignment; S35, at encoder output Graph-level convergence vector Access Domain Identification Subnet The subnet outputs domain labels in a binary classification format. By minimizing the adversarial loss source domain, the target domain confusion item is identified: ； in, To counteract domain alignment loss; , For graph-level embedding of the source and target domains; For domain discrimination subnet parameter set ; by The two-player game mode alternates and updates, driving the encoder to erase domain differences, while the discriminant subnet continuously distinguishes domain labels, achieving convergence of high-order feature distributions; S36. To suppress geometric dominance drift during the adversarial training phase, at each mapping node... Calculate statistical difference ,in , These are the batch averages for temperature and curvature, respectively; Weighted Inclusion This enables node-level statistical matching, prompting the temperature difference and morphological benchmark of the two domains to converge simultaneously. S37, When the accuracy of local discrimination drops to and At this point, adversarial training is stopped, all encoder weights are frozen, the source domain equalization parameters are mapped to the target domain feature map, and an alignment map with unified semantics is output. This is used for subsequent generation of adversarial graph transfer networks.

5. The method for predicting and optimizing warpage of complex carbon fiber composite cantilever structures according to claim 1, characterized in that, S4 includes the following steps: S41, in the alignment diagram The generator and discriminator are directly connected; the first layer of the generator constructs a fiber orientation adaptive graph convolution, which differs from a general adaptive graph convolution in that it is based on the node fiber vector. Dynamically modulate the original edge weights Obtain the right to amend : ； in, The basic edge weights are calculated in step S2; This is the fiber gain coefficient; For nodes Normalized fiber orientation vector; Connect the nodes with unit vectors ; S42. Training phase discriminator parameters locking Only high-fidelity finite element warp field The generator is trained through ten rounds of reconstruction; the loss term is a mixture of node mean squared error and local gradient difference to ensure the generator output... Quickly capture the overall warping trend and avoid gradient oscillations caused by early counter-offsets; S43. After preheating, the thawing detector sends... Synchronous input and The discriminator calculates the peak stress difference at the nodes, the second-order Laplace smoothness of the deformation field, and the strain energy difference across the entire graph, and outputs a domain confidence score. ; S44. During the alternating update process, the overall objective function is set as follows: ； in, This refers to the error in the reconstruction of the nodal field. To combat cross-entropy loss, drive confuse ; This represents the sum of squares of the peak stress differences; The residual is a second-order Laplace. This refers to the global strain energy closure error. For stage-adaptive weights, based on Dynamic adjustment; Through backpropagation of the above equation, when When outputting high confidence for three consecutive rounds, the residual Laplace channel weights are automatically increased to promote rapid convergence of details; S45. Real-time monitoring of the fiber parallel error curve; when the error descent gradient in this direction falls below a set threshold, immediately relax the weight decay and increase the learning rate. ; S46. After each iteration, calculate the energy closure error on the independent verification graph. ;like Exceeding the project limit That is, back up the generator to the previous stable weight, and... reduce ; S47, When the comprehensive error The decline was less than 5 times in five consecutive rounds. and At this point, freeze all parameters of the generator and discriminator, and export the trained generator. The generator provides a reliable initial deformation field for the multi-objective joint optimization in step S5, while ensuring node-level physical consistency.

6. The method for predicting and optimizing warpage of complex carbon fiber composite cantilever structures according to claim 1, characterized in that, S5 includes the following steps: S51, Reader / Discriminator The latest inference results yielded the node smoothness residual vectors respectively. Peak stress difference vector scalar of overall energy deviation The three types of physical indicators are concatenated into a physical consistency vector: ； This provides a quantitative benchmark for subsequent weight scheduling; in, The node's second-order Laplace smooth residual vector; The peak stress difference vector at the nodes; This represents the overall strain energy deviation. It is the vector norm 2; It is the infinite norm of a vector; S52, Output warp field in generator With high-fidelity samples Calculate node-level reconstruction error and the physical consistency vector of step S51 and counter-matching error Fusion, constructing a multi-objective loss function: ； in, The total loss; For adaptive weights; The total number of nodes; For the generator Nodal equivalent stress; To design the allowable stress threshold; This is the energy constraint scaling factor; The generator outputs an equivalent variable energy; This represents the theoretical minimum strain energy calculated according to engineering standards. S53, using a length of Sliding window record Historical mean and variance; batch normalization is used to scale the magnitude of each loss gradient to... This prevents a single component from dominating backpropagation and establishes a unified numerical scale for weighted scheduling. S54. Design a dynamic weight scheduler based on the physical consistency vector. Real-time changes and adjustments Specifically, it adopts an exponential ratio update: ； in, For the first Round Weight; This is the scheduling gain coefficient; For the first in the sliding window The latest mean squared error of the components; This is the total mean square error of the three components, used for normalization; When physical indicators exceed the threshold Increase, make This results in exponential amplification; if the visual reconstruction error remains high, then... The weights are increased accordingly, thereby dynamically balancing visual accuracy and physical plausibility during training; S55. Perform alternating optimization, freezing the discriminator in each outer iteration cycle and only using... Reverse update generator; if detected If the oscillation exceeds the set threshold, the reconstruction channel gradient will be temporarily frozen for two rounds to prevent the fiber principal direction details from being weakened due to excessive smoothing. S56, every other time The wheel is evaluated on an independent verification plot to assess the maximum nodal stress. Energy difference with the whole image When both are below the engineering tolerance line Then, lock the current weight range and lower the learning rate. The process has now entered the detailed refinement stage, with a focus on optimizing local warping peaks and microscale stress concentrations. S57, when continuous All rounds of verification are satisfied , and The rate of change is lower than At that point, training is terminated, and the generator's final weights are saved. With the dual guarantee of morphological consistency and physical reliability, the weights directly serve as a highly reliable initial field for subsequent local fine-tuning and process path optimization.

7. The method for predicting and optimizing warpage of complex carbon fiber composite cantilever structures according to claim 1, characterized in that, S6 includes the following steps: S61. Obtain the geometric-physical feature map of the structure to be predicted. Input freeze generator Obtained the first edition of the warp field Simultaneously calculate node gradients Equivalent stress estimation with anisotropy And cache it as a risk assessment benchmark; S62, Synthesize curvature jump variables for each node Predicted peak stress ratio Historical error replay item Construct a risk index: ； in To normalize the weights, for Sort by highest to lowest, and extract the first few. Nodes are aggregated into a high-risk subdomain set based on topological connectivity. ; S63. Process each subdomain sequentially. The local adaptive mesher is invoked to generate a refined mesh based on the nodal curvature gradient, and then a second-level low-fidelity heat transfer-elasticity solver is driven, with the same furnace temperature curve as input. Output local warp correction amount The correction results are mapped to the original image node indices using nearest neighbor interpolation. S64. Enable a writable channel in the graph network refinement layer. After writing back to the corresponding node, use a Gaussian kernel. The diffusion extends two layers of neighborhood along the graph edge, using a weight decay factor during the diffusion process. Ensure a smooth local-to-global transition; S65. Keep the backbone parameters frozen, only unlock the refinement layer weights. Select the difference before and after correction As a monitoring signal, based on mini-batch incremental gradient pairs conduct The network is updated step by step to add local features to its memory, thus avoiding global overfitting. S66. Obtain the second version of the warp field using updated network inference. If the residual error of all high-risk nodes Less than the threshold If the threshold scaling factor is correct, the final warp field will be output; otherwise, the threshold scaling factor will be adjusted. Multiply Then return to S62 and continue iterating until the accuracy requirements are met; S67. The image that has been fine-tuned. Local correction records With updated weights Unified archiving to the source domain sample pool; triggering main network retraining during daily offline phases, using the forgetting rate. The experience replay strategy continuously injects new local knowledge to achieve the continuous evolution of overall prediction accuracy.

Images