A method and system for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning
By integrating autoencoders and deep learning, a mechanical response prediction system for cellular materials was constructed, which solved the problem of predicting the mechanical response of cellular materials under complex loads. This system achieves efficient and accurate mechanical performance evaluation and is applicable to fields such as aerospace, rail transportation, and automotive collision safety protection.
Patent Information
- Application Number
- CN202611535094.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to efficiently and accurately predict the mechanical response of cellular materials under complex loads, especially under out-of-plane shear-compression combined loading conditions. Experimental methods are costly, theoretical analytical methods struggle to handle geometric irregularities, and finite element simulations are time-consuming, making it difficult to meet design and optimization requirements.
A cellular finite element model library is constructed by integrating autoencoders and deep learning. Through a three-dimensional convolutional autoencoder and a multi-task long short-term memory neural network, the force-displacement-energy time series and nodal displacement vectors of cellular materials are predicted to generate mechanical response results, which can adapt to unstructured mesh data and geometric irregularities.
It improves the accuracy and comprehensiveness of mechanical property evaluation of cellular materials, reduces simulation time, and can simultaneously predict force-displacement curves, energy absorption history and mesh deformation evolution, making it suitable for complex load analysis of practical cellular materials.
Smart Images

Figure CN123069798A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of material mechanical property analysis technology, specifically to a method and system for predicting the mechanical response of cellular materials that integrates autoencoders and deep learning. Background Technology
[0002] Honeycomb sandwich composites are widely used in aerospace, rail transportation, and automotive collision safety protection due to their lightweight, high specific strength, high specific stiffness, and excellent energy absorption capacity. During service, honeycomb materials typically withstand complex multiaxial loads, especially out-of-plane shear-compression combined loading conditions. Accurate prediction of their mechanical response is crucial for structural design and performance optimization.
[0003] Currently, obtaining the mechanical behavior of cellular materials mainly relies on three types of methods: (1) Experimental testing method: Force-displacement curves and deformation modes were obtained through quasi-static compression, dynamic impact and shear-compression combined loading experiments. The results of this method are reliable, but it has drawbacks such as high specimen preparation cost, long testing cycle and difficulty in obtaining full-field deformation information. Moreover, for the requirement of multi-condition parametric analysis, the experimental cost increases exponentially. In addition, the experimental method is difficult to separate the normal and shear response components under combined loading, which limits the depth of its verification of constitutive models.
[0004] (2) Theoretical analysis method: Based on the super folded element theory, the axial average crushing stress of the hexagonal honeycomb structure is analyzed. However, the analytical method usually relies on idealized geometric assumptions (such as perfect periodic hexagons and uniform wall thickness), which makes it difficult to deal with the geometric irregularities, wall thickness non-uniformity and material nonlinearity of actual honeycomb.
[0005] (3) Finite element numerical simulation method: By establishing a refined cellular finite element model through explicit dynamic software such as LS-DYNA and ABAQUS, the progressive folding, buckling and failure process of the structure can be accurately captured. However, the number of elements in the refined model can reach hundreds of thousands, and a single simulation takes several to tens of hours, which is difficult to meet the needs of design space exploration, parameter sensitivity analysis and multi-objective optimization for massive simulation data.
[0006] Therefore, there is an urgent need for a method that can accurately and efficiently predict the mechanical behavior of cellular materials under complex loads. Summary of the Invention
[0007] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for predicting the mechanical response of cellular materials that integrates autoencoders and deep learning.
[0008] According to one aspect of this application, a method for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning is provided, comprising: constructing a cellular finite element model library based on the geometric parameters of the cellular structure; wherein the cellular finite element model library contains multiple cellular finite element models corresponding to different combinations of geometric parameters; performing anisotropic mechanical feature voxelization processing on the cellular finite element models to generate multi-channel voxel maps; constructing and training a physically constrained three-dimensional convolutional autoencoder, and using the three-dimensional convolutional autoencoder to extract the low-dimensional latent space vector of the multi-channel voxel map; constructing and training an energy-conservation-constrained multi-task long short-term memory neural network, and using the low-dimensional latent space vector as the input of the multi-task long short-term memory neural network to predict the force-displacement-energy time series of the cellular finite element model; constructing and training a mesh deformation prediction network, and using the low-dimensional latent space vector, time step, and initial node coordinates as input to predict node displacement vectors; and generating the mechanical response result of the cellular finite element model based on the force-displacement-energy time series and the node displacement vectors.
[0009] In one embodiment, the anisotropic mechanical feature voxelization process of the cellular finite element model to generate a multi-channel voxel map includes: determining the overall spatial bounding box of the cellular finite element model, and mapping the node coordinates of the overall spatial bounding box to the voxel space based on a set voxel resolution to obtain voxel indices; calculating the lumped mass of each node, and accumulating the lumped mass to the corresponding voxel using a superposition method to obtain a basic mass voxel map; constructing an anisotropic mechanical feature tensor voxel map; wherein the anisotropic mechanical feature tensor voxel map includes: out-of-plane compressive stiffness contribution factor, first out-of-plane shear stiffness contribution factor, second out-of-plane shear stiffness contribution factor, local buckling strength factor, mass distribution correction factor, and initial geometric defect features; and concatenating the basic mass voxel map and the anisotropic mechanical feature tensor voxel map by channel splicing to form the multi-channel voxel map.
[0010] In one embodiment, constructing the anisotropic mechanical characteristic tensor voxel map includes: calculating the out-of-plane compressive stiffness contribution factor, the first out-of-plane shear stiffness contribution factor, the second out-of-plane shear stiffness contribution factor, the local buckling strength factor, the mass distribution correction factor, and the initial geometric defect characteristics based on the geometric parameters of the honeycomb structure, the out-of-plane equivalent modulus of the honeycomb, the out-of-plane shear modulus of the honeycomb, the substrate flow stress, the cell wall length, and the distribution of double-layer walls and single-layer walls.
[0011] In one embodiment, constructing and training the physically constrained 3D convolutional autoencoder includes: constructing an initial autoencoder; wherein the initial autoencoder includes an encoder network and a decoder network; constructing a physical constraint loss function based on channel reconstruction loss, physical consistency loss, and latent spatial regularization loss; and training the initial autoencoder based on the physical constraint loss function to obtain the 3D convolutional autoencoder.
[0012] In one embodiment, constructing and training the multi-task long short-term memory neural network with energy conservation constraints includes: constructing an initial neural network; wherein the initial neural network includes an LSTM layer sequence and a multi-task output layer; constructing a multi-task loss function for energy conservation constraints based on the mean square error of force, displacement, energy prediction, and energy conservation constraints; and training the initial neural network based on the multi-task loss function to obtain the multi-task long short-term memory neural network.
[0013] In one embodiment, constructing and training the mesh deformation prediction network includes: constructing an initial prediction network; the initial prediction network includes a multilayer perceptron architecture with residual connections; constructing a deformation prediction loss function based on smoothness constraints; and training the initial prediction network based on the deformation prediction loss function to obtain the mesh deformation prediction network.
[0014] In one embodiment, generating the mechanical response results of the cellular finite element model based on the force-displacement-energy time series and the node displacement vector includes: generating the force-displacement curve, energy absorption curve, and deformation configuration diagram at different times of the cellular finite element model based on the force-displacement-energy time series and the node displacement vector.
[0015] In one embodiment, the method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning further includes: acquiring experimental modal data and compressed experimental data of the cellular structure; calculating the sensitivity index of the experimental modal data to each dimension of the potential space; constructing a joint loss function based on the error between the experimental modal data, the compressed experimental data, and the simulation data; and fine-tuning the parameters of the three-dimensional convolutional autoencoder and the multi-task long short-term memory neural network based on the joint loss function and the sensitivity index.
[0016] In one embodiment, the step of fine-tuning the parameters of the three-dimensional convolutional autoencoder and the multi-task long short-term memory neural network based on the joint loss function and the sensitivity index includes: fixing the parameters of the first two layers of the three-dimensional convolutional autoencoder and the parameters of the multi-task long short-term memory neural network that are unrelated to the sensitivity index, and only updating the last two layers of the three-dimensional convolutional autoencoder, the connection weights of the multi-task long short-term memory neural network corresponding to the sensitivity index, and the output layer parameters.
[0017] According to another aspect of this application, a system for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning is provided, comprising: a model library construction module for constructing a cellular finite element model library based on the geometric parameters of the cellular structure; wherein the cellular finite element model library contains multiple cellular finite element models corresponding to different combinations of geometric parameters; a voxel map generation module for performing anisotropic mechanical feature voxelization processing on the cellular finite element models to generate multi-channel voxel maps; and an encoder construction module for constructing and training a physically constrained three-dimensional convolutional autoencoder, and using the three-dimensional convolutional autoencoder to extract the multi-channel voxel maps. The system includes: a low-dimensional latent space vector; a neural network construction module for constructing and training a multi-task long short-term memory neural network with energy conservation constraints, and using the low-dimensional latent space vector as input to predict the force-displacement-energy time series of the cellular finite element model; a deformation network construction module for constructing and training a mesh deformation prediction network, and using the low-dimensional latent space vector, time step, and initial node coordinates as input to predict node displacement vectors; and a response result generation module for generating the mechanical response results of the cellular finite element model based on the force-displacement-energy time series and the node displacement vectors.
[0018] This application provides a method and system for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning. Based on the geometric parameters of the cellular structure, a cellular finite element model library is constructed. This library contains multiple cellular finite element models corresponding to different combinations of geometric parameters. The cellular finite element models are anisotropically voxelized to generate multi-channel voxel maps. A physically constrained three-dimensional convolutional autoencoder is constructed and trained, and the low-dimensional latent space vector of the multi-channel voxel map is extracted using the three-dimensional convolutional autoencoder. An energy-conservation-constrained multi-task long short-term memory neural network is constructed and trained, and the low-dimensional latent space vector is used as the input to the multi-task long short-term memory neural network to predict the force-response relationship of the cellular finite element model. Displacement-energy time series; construct and train a mesh deformation prediction network, and predict node displacement vectors using low-dimensional latent space vectors, time steps, and initial node coordinates as inputs; generate mechanical response results of the cellular finite element model based on the force-displacement-energy time series and node displacement vectors; integrate the intrinsic mechanical information of cellular materials into voxel maps, enabling the deep learning model to learn the orientation-dependent characteristics of the material, and the voxelized feature extraction is adapted to unstructured mesh data, which can handle the geometric irregularities commonly found in actual cellular structures. At the same time, a multi-task learning architecture is adopted to simultaneously predict force-displacement curves, energy absorption history, and mesh deformation evolution, thereby improving the accuracy and comprehensiveness of the mechanical performance evaluation of cellular materials. Attached Figure Description
[0019] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.
[0020] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of a method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning.
[0021] Figure 2 This is a schematic diagram of the structure of a system for predicting the mechanical response of cellular materials that integrates autoencoders and deep learning, provided in an exemplary embodiment of this application.
[0022] Figure 3 This is a structural diagram of an electronic device provided in an exemplary embodiment of this application. Detailed Implementation
[0023] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.
[0024] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of a method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning. Figure 1 As shown, this method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning includes: Step 110: Construct a cellular finite element model library based on the geometric parameters of the cellular structure.
[0025] The cellular finite element model library contains multiple cellular finite element models corresponding to different combinations of geometric parameters. Geometric parameters include cellular height H, width B, length L, wall thickness h, branch angle α, and single / double layer wall distribution. This application generates N sets of geometric parameter combinations in the parameter space. For each combination, a refined finite element model of the cellular structure is established, using shell elements for mesh generation. Material constitutive models, boundary conditions (bottom fixed support, top rigid plate loaded with initial velocity), and loading conditions (including uniaxial compression, pure shear, and combined shear-compression loading) are set.
[0026] Step 120: Perform anisotropic mechanical feature voxelization processing on the cellular finite element model to generate a multi-channel voxel map.
[0027] Step 130: Construct and train a physically constrained 3D convolutional autoencoder, and use the 3D convolutional autoencoder to extract the low-dimensional latent space vector of the multi-channel voxel map.
[0028] Step 140: Construct and train a multi-task long short-term memory neural network with energy conservation constraints, and use a low-dimensional latent space vector as the input of the multi-task long short-term memory neural network to predict the force-displacement-energy time series of the cellular finite element model.
[0029] Step 150: Construct and train the mesh deformation prediction network, and predict the nodal displacement vectors using the low-dimensional potential space vector, time step, and initial node coordinates as inputs.
[0030] Step 160: Based on the force-displacement-energy time series and nodal displacement vectors, generate the mechanical response results of the cellular finite element model.
[0031] This application provides a method for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning. Based on the geometric parameters of the cellular structure, a cellular finite element model library is constructed. This library contains multiple cellular finite element models corresponding to different combinations of geometric parameters. The cellular finite element models are anisotropically voxelized to generate multi-channel voxel maps. A physically constrained 3D convolutional autoencoder is constructed and trained, and the low-dimensional latent space vector of the multi-channel voxel map is extracted using the 3D convolutional autoencoder. An energy-conservation constrained multi-task long short-term memory neural network is constructed and trained, and the low-dimensional latent space vector is used as the input to the multi-task long short-term memory neural network to predict the force-potential of the cellular finite element model. Force-displacement-energy time series; construct and train a mesh deformation prediction network, and predict node displacement vectors using low-dimensional latent space vectors, time steps, and initial node coordinates as inputs; generate mechanical response results of the cellular finite element model based on the force-displacement-energy time series and node displacement vectors; integrate the intrinsic mechanical information of cellular materials into voxel maps, enabling the deep learning model to learn the orientation-dependent characteristics of the material, and the voxelized feature extraction is adapted to unstructured mesh data, which can handle the geometric irregularities commonly found in actual cellular structures. At the same time, a multi-task learning architecture is adopted to simultaneously predict force-displacement curves, energy absorption history, and mesh deformation evolution, thereby improving the accuracy and comprehensiveness of the mechanical performance evaluation of cellular materials.
[0032] In one embodiment, step 120 can be implemented as follows: determine the overall spatial bounding box of the cellular finite element model, and map the node coordinates of the overall spatial bounding box to the voxel space based on the set voxel resolution to obtain the voxel index; calculate the lumped mass of each node, and use the superposition method to accumulate the lumped mass to the corresponding voxel to obtain the basic mass voxel map; construct an anisotropic mechanical characteristic tensor voxel map; wherein, the anisotropic mechanical characteristic tensor voxel map includes: out-of-plane compressive stiffness contribution factor, first out-of-plane shear stiffness contribution factor, second out-of-plane shear stiffness contribution factor, local buckling strength factor, mass distribution correction factor, and initial geometric defect features; and perform channel splicing between the basic mass voxel map and the anisotropic mechanical characteristic tensor voxel map to form a multi-channel voxel map.
[0033] Specifically, determine the overall spatial bounding box [x] of all cellular models. min x max ] × [y min y max ] ×[z min , z max ], where x min x max y min y max z min z max Let N be the minimum and maximum values in the three coordinate directions of the three-dimensional coordinate system, and let N be the voxel resolution. x × N y × N z (N is set after convergence test) x =20、N y =20、N z =100). For the i-th node of the k-th cellular model, its coordinates are (x... i y i , z i ), calculate the index of the voxel to which the node belongs: ; ; ; Among them, I i J i K i These represent the voxel values in the three coordinate directions of voxel space. This indicates rounding down to the nearest integer.
[0034] Calculate the lumped mass m of the i-th node i :m i =(1 / 4)ρ×A e ×he Where ρ is the material density, and A e h is the unit area. e The element thickness is used; the nodal mass is accumulated to the corresponding voxel using the stacking method to obtain the basic mass voxel map V. m (I,J,K). Then, an anisotropic mechanical characteristic tensor voxel diagram V is constructed. f (I,J,K,c), where channels c=1,...,6 correspond to: c=1 representing the out-of-plane compressive stiffness contribution factor, c=2 representing the first out-of-plane shear stiffness contribution factor (xz direction), c=3 representing the second out-of-plane shear stiffness contribution factor (yz direction), c=4 representing the local buckling strength factor based on the super-folded element theory, c=5 representing the mass distribution correction factor, and c=6 representing the initial geometric defect characteristics. Finally, the basic mass voxel map and the anisotropic mechanical characteristic tensor voxel map are concatenated to form a multi-channel voxel map V(I,J,K,7)∈ ^(20×20×100×7) is used as the input for subsequent deep learning models.
[0035] In one embodiment, the specific implementation of step 120 above may be as follows: based on the geometric parameters of the honeycomb structure, the out-of-plane equivalent modulus of the honeycomb, the out-of-plane shear modulus of the honeycomb, the substrate flow stress, the cell wall length, and the distribution of double-layer walls and single-layer walls, calculate the out-of-plane compressive stiffness contribution factor, the first out-of-plane shear stiffness contribution factor, the second out-of-plane shear stiffness contribution factor, the local buckling strength factor, the mass distribution correction factor, and the initial geometric defect characteristics.
[0036] Specifically, the out-of-plane compressive stiffness contribution factor κz=(E cz ·h) / H, where E cz The out-of-plane equivalent modulus of the honeycomb structure; the contribution factor κ of the first out-of-plane shear stiffness. xz =G cxz ·h / H, where G cxz The external shear modulus of the xy-plane of the honeycomb structure; the contribution factor κ to the external shear stiffness of the second plane. yz =G cyz ·h / H, G cyz The external shear modulus of the yz-plane of the honeycomb structure; the local buckling intensity factor κ based on the super-folded element theory. b =5σ0·(h / l ) / 3, where σ0 is the substrate flow stress (taken as 120MPa). l Cell wall length (10 mm); Mass distribution correction factor κ for the double-wall region. m The mass distribution correction factor κ for the single-layer wall region is assigned a value of 2.0. m Assign a value of 1.0; initial geometric defect feature κ dLinear interpolation is performed based on the preset defect amplitude (0.1 mm), ranging from 0.8 to 1.2.
[0037] In one embodiment, step 130 can be implemented as follows: constructing an initial autoencoder; wherein the initial autoencoder includes an encoder network and a decoder network; constructing a physical constraint loss function based on channel reconstruction loss, physical consistency loss and latent spatial regularization loss; training the initial autoencoder based on the physical constraint loss function to obtain a three-dimensional convolutional autoencoder.
[0038] The encoder network consists of four 3D convolutional layers with kernel sizes of 3×3×3, stride of 2, and channel numbers of 32, 64, 128, and 256 respectively. Each layer is followed by a batch normalization layer and a ReLU activation function. The output of the last layer is flattened and mapped to a latent space vector z∈ d d=32; The decoder network consists of 4 layers of 3D transposed convolutions, symmetrical to the encoder, with channel numbers of 256, 128, 64, and 7 respectively. The last layer uses the Sigmoid activation function to map the output value to the [0,1] interval. The output is restored to 20×20×100×7 through output padding / cropping, and the output is a reconstructed voxel map V. recon (I,J,K,7). This application constructs a physical constraint loss function based on channel reconstruction loss, physical consistency loss, and latent spatial regularization loss: L total =L recon +αL phy +βL reg ; Among them, L recon To determine the reconstruction loss, a weighted mean square error is used, L. phy The physical consistency loss is used to constrain the mapping relationship between the latent space vector and the macroscopic mechanical characteristic parameters, L. phy =‖f σ (z)-σ m ||²+||f λ (z)-λ‖², where f σ (z) and f λ (z) is a learnable mapping function (implemented using a two-layer fully connected network, with a hidden layer dimension of 64 and an output dimension of 1), σ m L represents the average crushing stress, λ is the folding wavelength; reg For the potential space regularization loss, L reg =‖z‖², α=0.1, β=0.01 are balance coefficients.
[0039] This application employs an optimizer to train the autoencoder end-to-end with an initial learning rate of 0.001 and 100,000 training epochs (early stopping mechanism: if the verification loss does not decrease for 100 consecutive epochs, the training is stopped). After training, the encoder network parameters are fixed for use in the latent space extraction of subsequent steps.
[0040] In one embodiment, the specific implementation of step 140 above may be as follows: constructing an initial neural network; wherein the initial neural network includes an LSTM layer sequence and a multi-task output layer; constructing a multi-task loss function with energy conservation constraints based on the mean square error of force, displacement, and energy prediction and energy conservation constraints; training the initial neural network based on the multi-task loss function to obtain a multi-task long short-term memory neural network.
[0041] The LSTM layer sequence consists of two stacked LSTM (Long Short-Term Memory) units, each containing 256 hidden units. For time step t, the input to the LSTM unit includes the latent space vector z (which remains unchanged throughout the time series and is copied to each time step via a broadcast mechanism) and the time-step normalized value τ. t =2(t-1) / (T max -1)-1∈[-1,1](T max =150, t=1,2,…,150); the multi-task output layer includes three parallel branches, each a 2-layer fully connected network (128-dimensional hidden layers, ReLU activation function), outputting: normalized force value. t ∈[-1,1], normalized shift value ∈[-1,1], normalized energy value ∈[-1,1]. This application constructs a multi-task loss function with energy conservation constraints based on the mean square error of force, displacement, and energy prediction, and the energy conservation constraint term: L LSTM =γ F MSE F +γ u MSE u +γ E MSE E +γ cons L cons ; Among them, force loss MSE F = Displacement loss Energy loss , Total duration , , These represent the force, displacement, and energy values at time step t, along with the energy conservation constraint term. ,in Furthermore, the energy conservation constraint is calculated in the physical quantity space or using a consistent normalization method, with the weighting coefficient γ. F =1.0、γ u =1.0、γ E =0.5、γ cons =0.2.
[0042] The LSTM network was trained using the time-truncation backpropagation algorithm with an initial learning rate of 0.001 and 10,000 training epochs. The Dropout rate of 0.2 and the L2 regularization coefficient of 0.0001 were used to prevent overfitting. The training set ratio test showed that only 16% of the model (329 models) was needed to achieve good prediction results.
[0043] In one embodiment, step 150 can be implemented as follows: constructing an initial prediction network; the initial prediction network includes a multilayer perceptron architecture with residual connections; constructing a deformation prediction loss function based on smoothness constraints; and training the initial prediction network based on the deformation prediction loss function to obtain a mesh deformation prediction network.
[0044] The network uses a latent space vector z (32-dimensional) and a time step τ. t The input consists of (1D) and (3D) initial node coordinates (x0, y0, z0), and the output is the node's displacement vector (u) at time t. x ,u y ,u z (3D). This network employs a multilayer perceptron architecture with residual connections, containing four hidden layers, each with 256 neurons. The activation function is Swish, with a residual connection added every two layers. The output layer uses linear activation. A deformation prediction loss function is constructed based on smoothness constraints. ; in, The total number of nodes. , L represents the actual displacement and the predicted displacement at time t, respectively. smooth To smooth the constraint term, This is used to penalize cases where the displacement gradient between adjacent nodes is too large, where E is the total number of elements. Let be the displacement gradient at time t. =0.05 is the smoothing coefficient.
[0045] This application employs a joint training strategy with a multi-task long short-term memory neural network. After each LSTM update, the mesh deformation prediction network is trained using the latent space vector z output by the current encoder. The optimizer uses an initial learning rate of 0.0005, a batch size of 1024 (node level), and is trained for 500 epochs.
[0046] In one embodiment, step 160 can be implemented by generating force-displacement curves, energy absorption curves, and deformation configuration diagrams at different times for a cellular finite element model based on the force-displacement-energy time series and nodal displacement vectors.
[0047] For a cellular structure, its finite element mesh data is input, voxelized, and multi-channel voxel maps are generated. The latent space vector is extracted by an encoder and input into an LSTM multi-task network to predict the force-displacement-energy time series. The latent space vector and node coordinates are input into a displacement prediction network to predict the displacement of each node at each time step, reconstruct the deformation process, and output the prediction results by combining the force-displacement-energy time series and node displacement vectors. The prediction results include: force-displacement curves, energy absorption curves, and deformation configuration diagrams at different times.
[0048] In one embodiment, the above-mentioned method for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning may further include: acquiring experimental modal data and compressed experimental data of the cellular structure; calculating the sensitivity index of the experimental modal data to each dimension of the potential space; constructing a joint loss function based on the error between the experimental modal data, compressed experimental data and simulation data; and fine-tuning the parameters of the three-dimensional convolutional autoencoder and the multi-task long short-term memory neural network based on the joint loss function and the sensitivity index.
[0049] This application obtains experimental modal data (the first four natural frequencies under free-free boundary conditions) and compression experimental data (force-displacement curves) of a cellular structure, and calculates the sensitivity indices of the experimental data to each dimension of the potential space, identifying the k dimensions (k=6) with the highest sensitivity, thus forming a set of sensitive dimensions. A joint experimental-simulation loss function is constructed: L hybrid =L LSTM +μL exp ; Among them, L exp The error is the prediction error of the experimental data (a weighted sum of modal frequency error and force-displacement curve error), with μ=0.5 as the weight value.
[0050] In one embodiment, the parameter fine-tuning method for the above-mentioned three-dimensional convolutional autoencoder and multi-task long short-term memory neural network may be: fixing the parameters of the first two layers of the three-dimensional convolutional autoencoder and the parameters in the multi-task long short-term memory neural network that are unrelated to the sensitivity index, and only updating the parameters of the last two layers of the three-dimensional convolutional autoencoder, the connection weights in the multi-task long short-term memory neural network corresponding to the sensitivity index, and the output layer parameters.
[0051] Fine-tuning was performed using a transfer learning strategy: the parameters of the first two layers of the encoder and the parameters in the LSTM network that were independent of the sensitivity dimension were fixed. Only the parameters of the last two layers of the encoder, the connection weights in the LSTM network corresponding to the sensitivity dimension, and the output layer were updated, and the learning rate was reduced to 0.0001. The model was trained for 100 epochs. After fine-tuning, the model's prediction accuracy for real physical behavior was significantly improved: the intrinsic frequency prediction error decreased from 6.17% to 2.13%, and the error of the force-displacement curve in the compression experiment decreased by 42%.
[0052] The technical solution of this application will be illustrated below with examples: (1) Constructing a cellular finite element model library A full factorial experimental design was adopted, with the following geometric parameter ranges: Cellular height H: 375, 400, 425, 450, 475, 500 mm (6 levels); Cellular width B: 60, 65, 70, 75, 80, 85, 90 mm (7 levels); Cellular length L: fixed at 100 mm (to ensure mesh regularity); Wall thickness h: 1.50, 1.75, 2.00, 2.25, 2.50, 2.75, 3.00 mm (7 levels); Branch angle α: 90°, 100°, 110°, 120°, 130°, 140°, 150°, 160° (8 levels); Single and double wall distribution: standard hexagonal cellular configuration, with double walls located on the cell wall perpendicular to the L direction; Total number of parameter combinations: 6×7×7×8=2352, after removing geometrically invalid combinations, 2058 models are actually generated.
[0053] Finite element modeling parameters: Element type: Shell element, element size 5.0 mm × 5.0 mm, 3 integration points in the thickness direction; Material model: Piecewise linear plastic model, base material is aluminum alloy, density ρ = 2.7 × 10⁻⁶ -3 g / mm 3 Elastic modulus E = 70 GPa, Poisson's ratio ν = 0.33, yield stress σ y =80MPa, tensile strength σ u =173MPa, hardening parameters were calibrated according to material tests; loading condition: axial compression, initial velocity 15.56 m / s, added mass 560 kg, total simulation time 15 ms; contact definition: automatic single-sided contact (self-contact), static friction coefficient 0.45, dynamic friction coefficient 0.40; automatic surface-to-surface contact between rigid plate and specimen, static friction coefficient 0.45, dynamic friction coefficient 0.40; output frequency: force-displacement data is output every 0.1 ms (150 time points in total), deformation data is output every 0.5 ms (30 time points in total); computational resource consumption: average calculation time per model 207 seconds (16 cores in parallel), total calculation time approximately 118 hours (approximately 5 days).
[0054] (2) Voxelization of anisotropic mechanical features The voxel resolution N was determined through convergence testing. x ×N y ×N z =20×20×100, corresponding to a physical size of approximately 5.0mm×4.5mm×5.0mm for each voxel, matching the finite element mesh size. A typical element A... e =25mm², h e Taking 2.0mm as an example, the mass m of the node is... i =(1 / 4)×2.7×10 - ³×25×2.0=0.03375g. The anisotropic characteristic tensor is calculated as follows by accumulating the values to the corresponding voxels using the superposition method: For h=2.0mm and H=400mm, l =10mm, Ecz≈9.34GPa, κ z =46.7MPa; for h=2.0mm, G cxz ≈1.16 GPa, κ xz =5.8MPa; σ0 is taken as 120MPa (considering strain hardening), κ b =40MPa; κ m The value is assigned as 2.0 for the double-walled region and 1.0 for the single-walled region; κ d Linear interpolation was performed based on a preset initial perturbation amplitude (0.1 mm), ranging from 0.8 to 1.2. The final constructed voxel map has dimensions of 20×20×100×7, and the data size of a single model voxel map is 2.8×10⁻⁶. 5 One floating-point number; the total data volume of all 2058 models is approximately 5.76 × 10⁻⁶. 8 One floating-point number; approximately 2.3GB when stored in float32 format, and approximately 1.8GB after compression using HDF5 format.
[0055] (3) Construct and train a physically constrained 3D convolutional autoencoder The encoder employs four layers of 3D convolutions with kernel sizes of 3×3×3, a stride of 2, and channel numbers of 32, 64, 128, and 256 respectively. Each layer is followed by batch normalization and ReLU. The output of the final layer is flattened and mapped to a latent space vector z with dimension d of 32 through a fully connected layer. The decoder consists of four layers of 3D transposed convolutions, symmetrical to the encoder, with channel numbers of 256, 128, 64, and 7 respectively. The final layer uses a sigmoid activation function. The mapping function f... σ (z) and f λ(z) A two-layer fully connected network was used, with 64 hidden layers and 1 output dimension. The loss function weights were determined using grid search: α=0.1, β=0.01. Channel weights were set to 1.0 for c=1~6 and 0.5 for c=7. Training parameters: Adam optimizer, initial learning rate 0.001, cosine annealing decay, batch size 32, 100,000 training epochs (early stopping mechanism: stopping if the validation loss does not decrease for 100 consecutive epochs). 1646 models were randomly partitioned for training, and 412 models were used for validation. Training was conducted in parallel using four V100 GPUs. Training results: Training set reconstruction error MSE = 4.47 × 10⁻⁶. -6 The validation set error MSE is 5.21 × 10⁻⁶. -6 Physical consistency loss L phy =0.023, training time approximately 10 hours.
[0056] (4) Construct and train an energy conservation-constrained multi-task long short-term memory neural network. The LSTM network uses two stacked layers, each with 256 hidden units, and a time step T=150. Each multi-task output branch is a two-layer fully connected network (128-dimensional hidden layers with ReLU activation), and the output layer uses linear activation. The loss function weights γ... F =1.0, γ u =1.0, γ E =0.5, γ cons =0.2, the training set ratio test showed that 0.16 (329 models) could achieve good results, and the remaining 1729 models were used as the test set. Training parameters: Adam optimizer, initial learning rate 0.001, cosine annealing decay, batch size 32, training for 10,000 epochs, dropout rate 0.2, L2 regularization coefficient 0.0001. Test results: force prediction error = 0.0023 (after normalization), displacement prediction error = 0.0019, energy prediction error = 0.0042, average energy absorption prediction error 3.8%, average peak force prediction error 5.2%.
[0057] (5) Construct and train a mesh deformation prediction network The network input is a latent space vector z (32-dimensional) and a time step τ. t(1D), initial node coordinates (3D), total 36 dimensions. Four hidden layers, 256 neurons per layer, Swish activation function, residual connections added every two layers, output displacement vector (3D). Loss function smoothing coefficient η=0.05, training parameters: Adam optimizer, initial learning rate 0.0005, batch size 1024 (node level), 500 training epochs, jointly trained with an LSTM network (after each LSTM update, the deformable network is trained using the latent space vector z output by the current encoder). Test results: average node displacement prediction error 1.87mm, maximum displacement error approximately 5.0mm, average relative error 2.3%.
[0058] (6) Experimental data fusion and verification Single-model prediction time: Voxelization processing 0.3 seconds, encoder inference 0.02 seconds, LSTM multi-task network inference 0.00027 seconds (150 time steps), deformation prediction network inference 0.4 seconds (5000 nodes), totaling approximately 0.72 seconds per model. Compared to finite element simulation (207 seconds per model), this represents a speedup of 287 times (including deformation prediction) to approximately 7.67 × 10⁻⁶. 5 Times (LSTM multi-task network inference only).
[0059] The following results were obtained by selecting 20 non-trained cellular models for validation: Force-displacement curve: average correlation coefficient R with finite element results 2 =0.967, peak force average error 4.8%, energy absorption average error 3.9%; Deformation mode: The average image similarity is 0.89, and the similarity of the last frame is 0.84. It successfully predicted the transformation of two folding modes (CM1 and CM2). The predicted transformation angle is 142°, which is close to the theoretical value of 140°. Anisotropic feature ablation experiment: After removing the anisotropic feature tensor channels, the prediction accuracy under shear-compression joint loading decreased by 35%, verifying the effectiveness of the core innovation.
[0060] Modal experiments were conducted on an aluminum alloy honeycomb sandwich panel under free-free boundary conditions to obtain the first four natural frequencies; simultaneously, compression experiments were performed to obtain force-displacement curves. L was constructed using the experimental data. expThe loss term, with a weight μ=0.5, identifies six potential dimensions (3rd, 5th, 8th, 12th, 18th, and 25th dimensions) most sensitive to experimental data based on sensitivity analysis. The parameters of the first two layers of the encoder and the connection weights in the LSTM network unrelated to the sensitive dimensions are fixed. Only the parameters of the last two layers of the encoder, the connection weights in the LSTM network corresponding to the sensitive dimensions, and the output layer parameters are updated. The learning rate is reduced to 0.0001, and training is performed for 100 epochs. Fine-tuning results: the intrinsic frequency prediction error decreased from 6.17% to 2.13%; the MSE of the force-displacement curve in the compression experiment decreased by 42%; and the identified out-of-plane shear modulus G... cxz =98.2MPa, which is in good agreement with the experimental results.
[0061] Figure 2 This is a schematic diagram of the structure of a system for predicting the mechanical response of cellular materials that integrates autoencoders and deep learning, provided in an exemplary embodiment of this application. Figure 2 As shown, the fusion autoencoder and deep learning cellular material mechanical response prediction system 20 includes: a model library construction module 21, used to construct a cellular finite element model library based on the geometric parameters of the cellular structure; wherein, the cellular finite element model library contains multiple cellular finite element models corresponding to different combinations of geometric parameters; a voxel map generation module 22, used to perform anisotropic mechanical feature voxelization processing on the cellular finite element models to generate multi-channel voxel maps; and an encoder construction module 23, used to construct and train a physically constrained three-dimensional convolutional autoencoder, and use the three-dimensional convolutional autoencoder to extract the low-dimensional potential of the multi-channel voxel map. In the spatial vector; the neural network construction module 24 is used to construct and train a multi-task long short-term memory neural network with energy conservation constraints, and uses the low-dimensional latent spatial vector as the input of the multi-task long short-term memory neural network to predict the force-displacement-energy time series of the cellular finite element model; the deformation network construction module 25 is used to construct and train a mesh deformation prediction network, and uses the low-dimensional latent spatial vector, time step and node initial coordinates as input to predict the node displacement vector; the response result generation module 26 is used to generate the mechanical response result of the cellular finite element model based on the force-displacement-energy time series and node displacement vector.
[0062] This application provides a system for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning. A model library construction module 21 constructs a cellular finite element model library based on the geometric parameters of the cellular structure; the library contains multiple cellular finite element models corresponding to different combinations of geometric parameters. A voxel map generation module 22 performs anisotropic mechanical feature voxelization processing on the cellular finite element models to generate multi-channel voxel maps. An encoder construction module 23 constructs and trains a physically constrained three-dimensional convolutional autoencoder and uses it to extract the low-dimensional latent space vector from the multi-channel voxel map. A neural network construction module 24 constructs and trains an energy-conservation-constrained multi-task long short-term memory neural network, using the low-dimensional latent space vector as input to predict the mechanical response of cellular materials. The finite element model's force-displacement-energy time series is used; the deformation network construction module 25 constructs and trains the mesh deformation prediction network, and predicts the node displacement vector using the low-dimensional latent space vector, time step, and initial node coordinates as input; the response result generation module 26 generates the mechanical response results of the cellular finite element model based on the force-displacement-energy time series and node displacement vectors; the mechanical intrinsic information of the cellular material is integrated into the voxel map, enabling the deep learning model to learn the orientation-dependent characteristics of the material, and the voxelized feature extraction is adapted to unstructured mesh data, which can handle the geometric irregularities commonly found in actual cellular structures. At the same time, a multi-task learning architecture is adopted to simultaneously predict the force-displacement curve, energy absorption history, and mesh deformation evolution, thereby improving the accuracy and comprehensiveness of the mechanical performance evaluation of cellular materials.
[0063] In one embodiment, the voxel map generation module 22 can be further configured to: determine the overall spatial bounding box of the cellular finite element model, and map the node coordinates of the overall spatial bounding box to the voxel space based on the set voxel resolution to obtain voxel indices; calculate the lumped mass of each node, and use the superposition method to accumulate the lumped mass to the corresponding voxel to obtain a basic mass voxel map; construct an anisotropic mechanical characteristic tensor voxel map; wherein, the anisotropic mechanical characteristic tensor voxel map includes: out-of-plane compressive stiffness contribution factor, first out-of-plane shear stiffness contribution factor, second out-of-plane shear stiffness contribution factor, local buckling strength factor, mass distribution correction factor, and initial geometric defect features; and perform channel splicing of the basic mass voxel map and the anisotropic mechanical characteristic tensor voxel map to form a multi-channel voxel map.
[0064] In one embodiment, the voxel generation module 22 described above can be further configured to: calculate, based on the geometric parameters of the honeycomb structure, the out-of-plane equivalent modulus of the honeycomb structure, the out-of-plane shear modulus of the honeycomb structure, the substrate flow stress, the cell wall length, and the distribution of double-layer walls and single-layer walls, the out-of-plane compressive stiffness contribution factor, the first out-of-plane shear stiffness contribution factor, the second out-of-plane shear stiffness contribution factor, the local buckling strength factor, the mass distribution correction factor, and the initial geometric defect characteristics.
[0065] In one embodiment, the encoder construction module 23 described above can be further configured to: construct an initial autoencoder; wherein the initial autoencoder includes an encoder network and a decoder network; construct a physical constraint loss function based on channel reconstruction loss, physical consistency loss and latent spatial regularization loss; and train the initial autoencoder based on the physical constraint loss function to obtain a three-dimensional convolutional autoencoder.
[0066] In one embodiment, the neural network construction module 24 can be further configured to: construct an initial neural network; wherein the initial neural network includes an LSTM layer sequence and a multi-task output layer; construct a multi-task loss function with energy conservation constraints based on the mean square error of force, displacement, and energy prediction and energy conservation constraints; and train the initial neural network based on the multi-task loss function to obtain a multi-task long short-term memory neural network.
[0067] In one embodiment, the aforementioned deformable network construction module 25 can be further configured to: construct an initial prediction network; the initial prediction network includes a multilayer perceptron architecture with residual connections; construct a deformable prediction loss function based on smoothing constraints; and train the initial prediction network based on the deformable prediction loss function to obtain a mesh deformable prediction network.
[0068] In one embodiment, the response result generation module 26 can be further configured to generate force-displacement curves, energy absorption curves, and deformation configuration diagrams at different times of the cellular finite element model based on the force-displacement-energy time series and nodal displacement vectors.
[0069] In one embodiment, the aforementioned fusion autoencoder and deep learning cellular material mechanical response prediction system 20 can be further configured to: acquire experimental modal data and compressed experimental data of the cellular structure; calculate the sensitivity index of the experimental modal data to each dimension of the potential space; construct a joint loss function based on the error between the experimental modal data, compressed experimental data and simulation data; and fine-tune the parameters of the three-dimensional convolutional autoencoder and the multi-task long short-term memory neural network based on the joint loss function and the sensitivity index.
[0070] In one embodiment, the above-mentioned fusion autoencoder and deep learning cellular material mechanical response prediction system 20 can be further configured to: fix the parameters of the first two layers of the three-dimensional convolutional autoencoder and the parameters in the multi-task long short-term memory neural network that are unrelated to the sensitivity index, and only update the parameters of the last two layers of the three-dimensional convolutional autoencoder, the connection weights in the multi-task long short-term memory neural network corresponding to the sensitivity index, and the output layer parameters.
[0071] Below, for reference Figure 3This application describes an electronic device according to embodiments thereof. The electronic device may be either or both of a first device and a second device, or a standalone device independent of them, which may communicate with the first device and the second device to receive acquired input signals from them.
[0072] Figure 3 A block diagram of an electronic device according to an embodiment of this application is illustrated.
[0073] like Figure 3 As shown, the electronic device 10 includes one or more processors 11 and memory 12.
[0074] The processor 11 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 10 to perform desired functions.
[0075] The memory 12 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 11 may execute the program instructions to implement the methods of the various embodiments of this application described above and / or other desired functions. Various contents such as input signals, signal components, and noise components may also be stored in the computer-readable storage medium.
[0076] In one example, the electronic device 10 may also include an input device 13 and an output device 14, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).
[0077] When the electronic device is a standalone device, the input device 13 can be a communication network connector for receiving the collected input signals from the first device and the second device.
[0078] In addition, the input device 13 may also include, for example, a keyboard, a mouse, etc.
[0079] The output device 14 can output various information to the outside, including determined distance information, direction information, etc. The output device 14 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.
[0080] Of course, for the sake of simplicity, Figure 3Only some of the components of the electronic device 10 relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device 10 may include any other suitable components depending on the specific application.
[0081] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of this application described in the "Exemplary Methods" section above.
[0082] The computer program product can be written in any combination of one or more programming languages to perform the operations of the embodiments of this application. The programming languages include object-oriented programming languages such as Java and C++, as well as conventional procedural programming languages such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.
[0083] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of this application described in the "Exemplary Methods" section above.
[0084] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.
[0085] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.
[0086] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.
[0087] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.
[0088] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.
[0089] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.
Claims
1. A method for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning, characterized in that, include: A cellular finite element model library is constructed based on the geometric parameters of the cellular structure; wherein, the cellular finite element model library contains multiple cellular finite element models corresponding to different combinations of geometric parameters; The cellular finite element model is subjected to anisotropic mechanical feature voxelization processing to generate a multi-channel voxel map; Construct and train a physically constrained 3D convolutional autoencoder, and use the 3D convolutional autoencoder to extract the low-dimensional latent space vector of the multi-channel voxel map; A multi-task long short-term memory neural network with energy conservation constraints is constructed and trained, and the low-dimensional latent space vector is used as the input of the multi-task long short-term memory neural network to predict the force-displacement-energy time series of the cellular finite element model. A mesh deformation prediction network is constructed and trained, and the node displacement vector is predicted using the low-dimensional potential space vector, time step and node initial coordinates as input. Based on the force-displacement-energy time series and the nodal displacement vector, the mechanical response results of the cellular finite element model are generated.
2. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 1, characterized in that, The process of anisotropically mechanically voxelizing the honeycomb finite element model to generate a multi-channel voxel map includes: The overall spatial bounding box of the cellular finite element model is determined, and the node coordinates of the overall spatial bounding box are mapped to the voxel space based on the set voxel resolution to obtain the voxel index. Calculate the lumped mass of each node, and use the superposition method to accumulate the lumped mass to the corresponding voxel to obtain the basic mass voxel map; Construct an anisotropic mechanical characteristic tensor voxel diagram; wherein, the anisotropic mechanical characteristic tensor voxel diagram includes: out-of-plane compressive stiffness contribution factor, first out-of-plane shear stiffness contribution factor, second out-of-plane shear stiffness contribution factor, local buckling strength factor, mass distribution correction factor, and initial geometric defect characteristics. The basic mass voxel map and the anisotropic mechanical characteristic tensor voxel map are spliced together via channels to form the multi-channel voxel map.
3. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 2, characterized in that, The construction of the anisotropic mechanical feature tensor voxel map includes: Based on the geometric parameters of the honeycomb structure, the out-of-plane equivalent modulus of the honeycomb, the out-of-plane shear modulus of the honeycomb, the substrate flow stress, the cell wall length, and the distribution of double-layer walls and single-layer walls, the out-of-plane compressive stiffness contribution factor, the first out-of-plane shear stiffness contribution factor, the second out-of-plane shear stiffness contribution factor, the local buckling strength factor, the mass distribution correction factor, and the initial geometric defect characteristics are calculated.
4. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 1, characterized in that, The construction and training of the physically constrained 3D convolutional autoencoder includes: Construct an initial autoencoder; wherein the initial autoencoder includes an encoder network and a decoder network; Based on channel reconstruction loss, physical consistency loss, and potential space regularization loss, a physical constraint loss function is constructed. The initial autoencoder is trained based on the physical constraint loss function to obtain the three-dimensional convolutional autoencoder.
5. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 1, characterized in that, The construction and training of the energy conservation-constrained multi-task long short-term memory neural network includes: Construct an initial neural network; wherein the initial neural network includes a sequence of LSTM layers and a multi-task output layer; Based on the mean square error of force, displacement, and energy predictions and the energy conservation constraint term, a multi-task loss function with energy conservation constraints is constructed. The initial neural network is trained based on the multi-task loss function to obtain the multi-task long short-term memory neural network.
6. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 1, characterized in that, The construction and training of the mesh deformation prediction network includes: Construct an initial prediction network; the initial prediction network includes a multilayer perceptron architecture with residual connections; Based on smoothness constraints, a deformation prediction loss function is constructed; The initial prediction network is trained based on the deformation prediction loss function to obtain the mesh deformation prediction network.
7. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 1, characterized in that, The mechanical response results of the cellular finite element model generated based on the force-displacement-energy time series and the nodal displacement vectors include: Based on the force-displacement-energy time series and the node displacement vector, the force-displacement curve, energy absorption curve, and deformation configuration diagram at different times of the cellular finite element model are generated.
8. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 1, characterized in that, The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning also includes: Obtain experimental modal data and compression experimental data of the cellular structure; Calculate the sensitivity index of the experimental modal data to each dimension of the potential space; Based on the errors between the experimental modal data, the compressed experimental data, and the simulation data, a joint loss function is constructed. Based on the joint loss function and the sensitivity index, the parameters of the three-dimensional convolutional autoencoder and the multi-task long short-term memory neural network are fine-tuned.
9. The method for predicting the mechanical response of cellular materials by fusing autoencoders and deep learning according to claim 8, characterized in that, The fine-tuning of the parameters of the 3D convolutional autoencoder and the multi-task long short-term memory neural network based on the joint loss function and the sensitivity index includes: The parameters of the first two layers of the three-dimensional convolutional autoencoder and the parameters in the multi-task long short-term memory neural network that are unrelated to the sensitivity index are fixed. Only the parameters of the last two layers of the three-dimensional convolutional autoencoder, the connection weights in the multi-task long short-term memory neural network corresponding to the sensitivity index, and the output layer parameters are updated.
10. A system for predicting the mechanical response of cellular materials by integrating autoencoders and deep learning, characterized in that, include: The model library construction module is used to construct a cellular finite element model library based on the geometric parameters of the cellular structure; wherein, the cellular finite element model library contains multiple cellular finite element models corresponding to different combinations of geometric parameters; The voxel generation module is used to perform anisotropic mechanical feature voxelization processing on the honeycomb finite element model to generate a multi-channel voxel map. The encoder construction module is used to construct and train a physically constrained 3D convolutional autoencoder, and to use the 3D convolutional autoencoder to extract the low-dimensional latent space vector of the multi-channel voxel map. The neural network construction module is used to construct and train a multi-task long short-term memory neural network with energy conservation constraints, and to predict the force-displacement-energy time series of the cellular finite element model using the low-dimensional latent space vector as the input of the multi-task long short-term memory neural network. The deformation network construction module is used to construct and train the mesh deformation prediction network, and predict the node displacement vector using the low-dimensional potential space vector, time step and node initial coordinates as input. The response result generation module is used to generate the mechanical response results of the cellular finite element model based on the force-displacement-energy time series and the nodal displacement vector.
