A design method of finite element model of mechanical properties of aluminum honeycomb core

Abstract

The application discloses a kind of design methods of aluminum honeycomb core mechanical property finite element model, it is related to material mechanics performance test technical field;In the application, first, construct multiscale defect coupling modeling module, extract manufacturing defect characteristics, generate multiscale model containing defect in combination with improved Voronoi algorithm and cohesive force model;Then establish bayesian optimization dynamic parameter correction module, with defect model data as sample, realize efficient multi-parameter optimization by proxy model and acquisition function;Subsequently, design LSTM drive dynamic damage prediction module, train model based on time series data and realize real-time damage early warning;Finally, build closed-loop collaborative architecture, realize two-way feedback and iterative optimization of each module by data interaction.The application improves model accuracy and design efficiency, realizes dynamic damage early warning, and is suitable for engineering design of aluminum honeycomb core lightweight structure.

Description

Technical Field

[0001] This invention relates to the field of material mechanical property testing technology, specifically to a design method for a finite element model of the mechanical properties of an aluminum honeycomb core. Background Technology

[0002] Aluminum honeycomb cores are widely used in key areas such as aerospace collision avoidance structures and new energy vehicle body buffer components due to their advantages of lightweight, high strength, and high energy absorption. In practical engineering applications, it is necessary to accurately predict the mechanical response of aluminum honeycomb cores under static and dynamic loads (such as compression and impact) using finite element models to guide structural design and performance optimization. Therefore, constructing a high-precision finite element model of the mechanical properties of aluminum honeycomb cores has become one of the core requirements of the industry.

[0003] Existing finite element method (FEM) models for the mechanical properties of aluminum honeycomb cores have significant limitations: some methods assume the aluminum honeycomb core is a "defect-free ideal structure," failing to account for unavoidable defects during manufacturing, such as cell wall loss and adhesive debonding, leading to large discrepancies between model predictions and actual experimental values; while some methods introduce defect modeling, parameter optimization relies on manual trial and error, resulting in low efficiency and a tendency to get trapped in local optima; still others can only predict static performance, failing to capture the damage evolution process under dynamic loads and lacking real-time early warning capabilities. In summary, the core problem with existing technologies is that they cannot simultaneously ensure physical realism including manufacturing defects, high efficiency in parameter optimization, and damage early warning capabilities under dynamic loads, resulting in long model design cycles, low reliability, and difficulty in meeting practical engineering needs.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] The purpose of this invention is to provide a design method for a finite element model of the mechanical properties of aluminum honeycomb cores, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides a design method for a finite element model of the mechanical properties of an aluminum honeycomb core, comprising the following steps:

[0007] Step 1: Construct a multi-scale defect coupling modeling module to extract manufacturing defect features of aluminum honeycomb cores and generate a multi-scale geometric model containing defects, and output the initial geometric model and basic mechanical dataset;

[0008] Step 2: Based on the basic mechanics dataset, construct a Bayesian optimization dynamic parameter correction module, and generate dynamic simulation data by iteratively searching for the optimal parameter combination through a surrogate model.

[0009] Step 3: Using dynamic simulation data as input, construct an LSTM-driven dynamic damage prediction module, train the model, and realize real-time damage state prediction.

[0010] Step four involves establishing a closed-loop collaborative architecture through a data interaction interface, comprising a multi-scale defect coupling modeling module, a Bayesian optimization dynamic parameter correction module, and an LSTM-driven dynamic damage prediction module. The damage warning signal output by the LSTM-driven dynamic damage prediction module serves as a dynamic constraint for the Bayesian optimization dynamic parameter correction module. The parameter combination output by the Bayesian optimization dynamic parameter correction module drives the multi-scale defect coupling modeling module to update the model, iterating until preset conditions are met before outputting the final finite element model of the aluminum honeycomb core's mechanical properties. By constructing a closed-loop collaborative architecture of "defect modeling - optimization - damage prediction," the limitations of independent operation of various modeling methods in existing technologies are overcome, achieving deep collaboration among multiple technologies. This significantly improves the accuracy and efficiency of the finite element model in predicting the mechanical properties of defective aluminum honeycomb cores, while also providing dynamic damage warning capabilities, thus offering more comprehensive theoretical support for aluminum honeycomb core structure design.

[0011] Furthermore, in step one, the construction process of the multi-scale defect coupling modeling module includes: acquiring manufacturing defect features of the aluminum honeycomb core using microscopic imaging technology, including cell wall defects, burrs, and bonding defects; generating a microscopic cell model containing manufacturing defect features based on the improved Voronoi algorithm, and introducing a cohesive force model to define crack propagation criteria at the defects; mapping the defect features of the microscopic cell model to the macroscopic finite element model through sub-model technology to obtain the initial geometric model; performing static compression simulation on the initial geometric model to output a basic mechanical dataset, which includes static compressive stress-strain curves under defect-free and defect-containing states; and constructing a multi-scale model by accurately extracting manufacturing defect features and using the improved Voronoi algorithm and cohesive force model to make the generated initial geometric model more closely match the actual structure of the aluminum honeycomb core, and the output basic mechanical dataset can truly reflect the influence of defects on mechanical properties, providing reliable physical basis data for subsequent parameter optimization and damage prediction.

[0012] Furthermore, in step two, the construction process of the Bayesian optimization dynamic parameter correction module includes: constructing a Gaussian process regression surrogate model using a basic mechanics dataset as samples; defining the optimization objective as maximizing platform stress and minimizing damage rate, determining the range of parameters to be optimized, including cell wall thickness and pore size distribution; iteratively searching for the optimal combination of parameters to be optimized using an expectation-improved acquisition function, and inputting the output combination of parameters to be optimized into a multi-scale defect coupling modeling module in each iteration to generate dynamic simulation data under the corresponding parameters, including stress response data under impact load; by constructing a Gaussian process regression surrogate model and using an expectation-improved acquisition function, the inefficiency of traditional trial-and-error methods is avoided, and the optimal parameter combination that balances platform stress and damage rate is quickly searched, while generating a large amount of high-quality dynamic simulation data to provide sufficient samples for LSTM model training and improve the accuracy of subsequent damage prediction.

[0013] Furthermore, in step three, the construction process of the LSTM-driven dynamic damage prediction module includes: labeling the dynamic simulation data in a time-series format to form a training dataset for the LSTM model, which includes defect parameters, load history, and damage evolution data; training the LSTM network based on the training dataset to enable the LSTM network to capture the nonlinear correlation between defect parameters, load history, and damage evolution; calibrating the trained LSTM network using strain gauge data obtained from drop hammer impact tests to ensure that the damage prediction error of the LSTM network meets preset requirements; embedding the calibrated LSTM network into the finite element simulation process to output the damage state prediction results in real time, which include crack propagation rate and remaining load-bearing capacity; and calibrating the time-series labeled training dataset with experimental data to enable the LSTM network to accurately capture the nonlinear relationship between the mechanical properties of aluminum honeycomb cores and damage evolution, thereby achieving real-time prediction of dynamic damage state and providing timely damage warning signals for subsequent parameter optimization, thus avoiding structural failure caused by optimized parameter combinations.

[0014] Furthermore, in step four, the process of establishing the data interaction interface includes: encapsulating the defect rate and cell wall thickness deviation parameters output by the multi-scale defect coupling modeling module into the initial parameter constraint range of the Bayesian optimization dynamic parameter correction module; pushing the parameter-mechanical response-damage state data generated by the Bayesian optimization dynamic parameter correction module to the training dataset of the LSTM-driven dynamic damage prediction module; when the damage rate predicted by the LSTM-driven dynamic damage prediction module exceeds a preset threshold, generating an emergency parameter correction signal and converting the emergency parameter correction signal into the dynamic constraint condition of the Bayesian optimization dynamic parameter correction module; through clear data interaction rules, the effective transmission and constraint transformation of key parameters among the three modules are realized, ensuring the smooth operation of the closed-loop collaborative architecture, enabling each module to adjust its own operating state in a timely manner according to the output of other modules, and avoiding the problem of decreased modeling accuracy caused by data disconnection between modules.

[0015] Furthermore, in step four, iterating until the preset conditions are met specifically involves: setting an iteration period and a prediction error threshold, and recording the mechanical property prediction error of the final finite element model in each iteration; when the mechanical property prediction error of three consecutive iterations is less than the prediction error threshold, and the damage warning signal output by the LSTM-driven dynamic damage prediction module is stable, the iteration is stopped and the final aluminum honeycomb core mechanical property finite element model is output; by setting a clear iteration termination condition, time wasted due to excessive iteration is avoided, while ensuring that the output final finite element model has stable mechanical property prediction accuracy and damage warning capability, thus guaranteeing the reliability and applicability of the model in practical engineering applications.

[0016] Furthermore, the improvement of the Voronoi algorithm lies in introducing actual manufacturing process parameters of aluminum honeycomb cores to constrain the generation boundary of the Voronoi diagram. These actual manufacturing process parameters include cell wall forming error and cutting accuracy parameters. By combining the actual manufacturing process parameters to improve the Voronoi algorithm, the generated micro-cell model is made more consistent with the actual production situation of aluminum honeycomb cores, further improving the authenticity of the output data of the multi-scale defect coupling modeling module and providing more accurate initial input for the operation of subsequent modules.

[0017] Furthermore, the parameters to be optimized also include the cell arrangement of the aluminum honeycomb core. During the iterative search process, the Bayesian optimization dynamic parameter correction module synchronously records the differences in dynamic simulation data under different cell arrangements and feeds the difference data back to the multi-scale defect coupling modeling module to optimize the arrangement structure of the micro-cell model. By incorporating the cell arrangement into the parameters to be optimized and feeding back the simulation difference data of different arrangements, the optimized parameter combination becomes more comprehensive. At the same time, it promotes the optimization of the micro-cell model arrangement structure of the multi-scale defect coupling modeling module, further improving the comprehensiveness and accuracy of the final finite element model in predicting the mechanical properties of the aluminum honeycomb core.

[0018] Furthermore, when calibrating the LSTM network using a drop hammer impact test, strain data is collected by attaching strain gauges at different locations on the aluminum honeycomb core. This strain data is then converted into damage state evaluation indicators and compared with the prediction results of the LSTM network. The weight parameters of the LSTM network are adjusted to reduce prediction errors. By collecting data from strain gauges at multiple locations and converting it into damage state evaluation indicators, the calibration process of the LSTM network becomes more accurate. This effectively corrects network prediction biases and further improves the damage prediction accuracy of the LSTM-driven dynamic damage prediction module, providing more reliable damage early warning support for the closed-loop collaborative architecture.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. By leveraging multi-scale defect coupling modeling, this approach overcomes the limitations of existing models that either ignore manufacturing defects or simplify only single defects. It extracts real manufacturing defect features, constructs a microscopic model containing multiple defect types using an improved Voronoi algorithm, introduces a cohesive force model to describe crack propagation, and then achieves multi-scale mapping through sub-model technology. This enables the model to realistically reproduce the actual structural state of the aluminum honeycomb core, significantly improving its physical realism and avoiding prediction biases caused by modeling based on ideal structures. This provides a reliable foundation for subsequent parameter optimization and damage prediction.

[0021] 2. By relying on Bayesian optimization for dynamic parameter correction, this approach addresses the problems of low efficiency and susceptibility to local optima in traditional parameter optimization. Using real data from multi-scale defect models as samples, a Gaussian process regression surrogate model is constructed. This model improves the balance of parameters in the acquisition function by improving expectations, while also incorporating cell arrangement into the multi-parameter optimization scope. This significantly shortens the parameter optimization cycle, eliminates the need for manual trial and error, and substantially improves optimization efficiency and effectiveness.

[0022] 3. By leveraging LSTM-driven dynamic damage prediction technology, this method overcomes the limitation of existing techniques in capturing nonlinear damage evolution under dynamic loads. Using time-series data generated through Bayesian optimization as training samples, a multi-layer LSTM network is designed to capture the correlation between defects, loads, and damage. The model is then calibrated through drop-weight impact tests to achieve real-time prediction and early warning of dynamic damage states. This technology not only improves the accuracy of dynamic damage prediction but also shortens the prediction response time, enabling early damage warnings and calculation of remaining load-bearing capacity, effectively avoiding false or missed warnings, and enhancing the model's engineering practical value.

[0023] 4. By constructing a closed-loop collaborative architecture of "defect modeling - optimization - damage prediction," this integrated solution breaks through the limitations of existing technologies where each module operates independently and data is transmitted unidirectionally. This architecture achieves bidirectional feedback between modules through a data interaction interface. It dynamically adjusts the parameter constraints of Bayesian optimization using the damage warning signal output by the LSTM, then drives the multi-scale model to update in real time, forming an iterative optimization closed loop. This not only significantly shortens the overall model design cycle but also further improves the reliability of the final model. It can be directly applied to the design of aluminum honeycomb core structures in aerospace, new energy vehicles, and other fields, significantly improving the design qualification rate and engineering adaptability. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the design principle of a finite element model for the mechanical properties of aluminum honeycomb cores. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1This invention provides a technical solution: a design method for a finite element model of the mechanical properties of aluminum honeycomb cores. This embodiment addresses the mechanical performance design requirements of aluminum honeycomb core collision-resistant structures in the aerospace field. In this scenario, the aluminum honeycomb core needs to withstand impact loads such as spacecraft landing cushioning and collisions with internal components of the cabin. Furthermore, issues such as missing cell walls and bonding defects are unavoidable during the manufacturing process. Traditional finite element models, due to their neglect of defects or low modeling efficiency, cannot meet the engineering requirements of high-precision prediction and rapid design.

[0027] I. Implementation Background: This implementation method takes an anti-collision panel with aluminum honeycomb core model 5052-H32 as the research object. The basic parameters of this aluminum honeycomb core are: cell type is regular hexagon, nominal cell wall thickness is... Nominal aperture Panel thickness The overall dimensions are length Width .

[0028] II. Construction of a Multi-Scale Defect Coupling Modeling Module: Existing technologies often assume no defects or simplify only a single defect in aluminum honeycomb core finite element modeling, such as the homogenized equivalent model described in the literature. However, in actual manufacturing processes, defects such as cell wall forming errors, cutting burrs, and adhesive debonding can cause the model's predicted values ​​to deviate from experimental values ​​by more than 12%. Therefore, it is necessary to construct a multi-scale model that can realistically reproduce defect characteristics, providing physically accurate initial input for subsequent parameter optimization and damage prediction. Specific technical methods are as follows:

[0029] 2.1 Collection and Quantification of Manufacturing Defect Features:

[0030] 1. Defect Acquisition Equipment and Methods: A super depth-of-field microscope was used to perform a full-area scan of the 5052-H32 aluminum honeycomb core sample. The scanning resolution was set to [missing information]. The scanning range covers the entire sample surface and cross-section; simultaneously, an ultrasonic flaw detector is used to detect internal bonding defects, with the probe frequency set to [value missing]. .

[0031] 2. Defect Quantification and Symbol Definition: Defining the Manufacturing Defect Rate That is, the proportion of defective cells out of the total number of cells. Define cell wall thickness deviation That is, the actual cell wall thickness With nominal thickness The difference, The unit is Define the ratio of adhesive bonding defect areas. This refers to the ratio of the area of ​​the adhesive-debonded region to the total adhesive area of ​​the cell wall. Define burr height This refers to the height of the protrusion at the edge of the cell wall during the cutting process, measured in units of... .

[0032] 3. Defect Statistical Results: Statistical analysis of the test data from three parallel samples yielded the defect distribution of this batch of aluminum honeycomb cores: , , , .

[0033] 2.2 Improved Microscopic Cell Model Construction of Voronoi Algorithm: Existing publicly available Voronoi algorithms only generate cells based on random seed points, without considering manufacturing process constraints. This results in large deviations between the generated cell shape and the actual shape, such as cell wall angle errors exceeding [a certain value]. .

[0034] The core design of the improved Voronoi algorithm:

[0035] Introducing manufacturing process parameter constraints: Including cell wall forming errors Cutting precision As a boundary condition for the algorithm, the cell side length is limited. The fluctuation range is cell wall angle ;

[0036] Embedding of defect features: for randomly selected Proportional cell units, by modifying the edge weights of the Voronoi diagram, simulate cell wall loss by deleting corresponding edges and adding spurs at the endpoints of the edges with a length of [missing information]. Line segments and adhesive detachment at cell wall junctions create dummy connections, which reduces the elastic modulus of that region.

[0037] Output of the microcellular model: Generates a microscopic model containing 1000 cells, with a model size of [size missing]. The output file is in ABAQUS .inp format.

[0038] 2.3 Crack propagation criterion definition of the cohesive force model (CZM): The existing publicly available Hashin failure criterion is only applicable to fiber-reinforced materials, and its prediction error for cell wall crack propagation in aluminum honeycomb cores exceeds 15%. The cohesive force model can describe the interface failure process at the defect through force-displacement curves, which is more in line with the cell wall fracture characteristics of aluminum honeycomb cores.

[0039] Cohesive model parameter settings: A bilinear cohesive model is adopted, whose constitutive relation satisfies:

[0040] ;

[0041] in, The maximum interfacial bond strength is defined as a value of [value missing]. ; For the linear segment displacement, the value is... ; The failure displacement is given by a value of .

[0042] The cohesive unit is embedded into the defects of the micro-cell model, such as the edge of the missing cell wall or the area of ​​adhesive debonding, and the unit type is set to COH3D8.

[0043] 2.4 Multiscale mapping and macroscopic model construction:

[0044] Mapping logic of sub-model technology:

[0045] Define the mesh size of the macroscopic model as Sub-model boundaries are set in the macro-region corresponding to the micro-model, and the boundary conditions are transmitted by displacement interpolation. That is, the nodal displacements of the macro-model are mapped to the boundary displacements of the micro-model through the interpolation algorithm.

[0046] Define the material parameters of the macroscopic model: the equivalent elastic modulus of the microscopic model. Equivalent Poisson ratio As homogenized material parameters in the macroscopic model, The results, obtained through unidirectional compression simulation of the microscopic model, satisfy the following: In the formula, The average stress of the microscopic model, The average strain.

[0047] 2. Output of the macroscopic model: Generates a size of The macroscopic finite element model, including a global mesh and sub-model regions, is output in ABAQUS .inp format. It also outputs a fundamental mechanics dataset, including static compressive stress-strain curves for the defect-free macroscopic model and the defective macroscopic model, with a data sampling frequency of [missing information]. .

[0048] Example: Multi-scale defect modeling of aerospace collision-resistant aluminum honeycomb cores: Input conditions: 5052-H32 aluminum honeycomb core sample, defect parameters , , , Static compressive load Loading speed Modeling process: The sample was scanned using a super depth-of-field microscope to obtain 8% of the defect cell locations. These cell locations were then designated as defect cells in the improved Voronoi algorithm and embedded... Burrs were removed; COH3D8 elements were added to the adhesive debonding area, and bilinear cohesive force model parameters were set; a macroscopic model was established, and a sub-model was set in the central area of ​​the anti-collision panel, which was associated with the microscopic defect model; static compression simulation was performed to obtain the stress-strain curve of the defect model: yield stress Platform stress Crushing strain Output: The basic mechanics dataset contains two curves, each with 1000 data points; the initial geometric model file size is approximately 500MB.

[0049] Existing publicly available multi-scale modeling methods only use the traditional Voronoi algorithm to generate defect-free microscopic models, without considering manufacturing process parameter constraints and real defect characteristics. Existing publicly available defect modeling only simulates adhesive bonding defects individually, failing to couple multiple defect types such as cell wall loss and burrs, and does not use a cohesive force model to describe crack propagation. The advantages of this approach are: improved modeling accuracy: the prediction error of static compressive stress in defect-containing models is reduced, and the prediction deviation of platform stress is minimized; improved data realism: the basic mechanics dataset includes the influence of real defects, providing samples of non-ideal working conditions for subsequent Bayesian optimization, avoiding engineering failures caused by optimization based on ideal models; improved process adaptability: the improved Voronoi algorithm, combined with manufacturing process parameters, generates a model that can be directly correlated with the actual production process, providing a quantitative basis for process optimization.

[0050] III. Construction of a Bayesian Optimization Dynamic Parameter Correction Module: In existing technologies, parameter optimization of aluminum honeycomb cores, such as cell wall thickness and pore size, relies on trial-and-error methods or single-objective optimization, such as pursuing only the maximization of plateau stress. This leads to long optimization cycles and a tendency to get trapped in local optima, such as high plateau stress but rapid damage rate. Therefore, it is necessary to construct a dynamic parameter correction module based on Bayesian optimization, using the basic data of a multi-scale defect model as samples to achieve multi-objective and efficient parameter optimization. Specific technical means are as follows:

[0051] 3.1 Definition of parameters to be optimized and optimization objectives:

[0052] 1. Definition of the parameter vector to be optimized and its symbol: Define the parameter vector to be optimized. ,in: Cell wall thickness, unit value range ; Cell pore size, unit value range ; The cell arrangement is a regular hexagonal arrangement. Rhomboid arrangement .

[0053] 2. Construction of the multi-objective optimization function: Define optimization objective 1: maximize platform stress, objective function ; Define optimization objective 2: minimize damage rate, objective function Define the weighted synthesis objective function: Among them, weight , ,satisfy .

[0054] 3.2 Construction of Gaussian Process Regression (GPR) surrogate model: The existing response surface model and neural network model described in public documents have fitting errors of more than 8% for nonlinear data, while Gaussian process regression can flexibly describe nonlinear relationships through kernel functions, with fitting errors of less than 3%, making it more suitable for modeling the mechanical response of aluminum honeycomb cores.

[0055] The core formula of the 1GPR model:

[0056] Define the output of the GPR model as ,in The error is random and follows a set pattern. ;

[0057] Define the kernel function as a quadratic exponential kernel that satisfies:

[0058] ;

[0059] in, For signal variance, For length scale, The Kronecker function;

[0060] Given a training sample set Then the new sample Predicted mean and prediction variance satisfy: ;

[0061] ;in, Let be the covariance vector between the new sample and the training samples. The covariance matrix of the training samples, It is an identity matrix.

[0062] 3.2 Optimization logic of the expected improvement (EI) acquisition function: The greedy algorithm described in existing public documents is prone to getting trapped in local optima and has low efficiency in random search; the expected improvement function can balance exploration and utilization, thereby improving the optimization efficiency.

[0063] 1. The core formula of the EI function: Definition For new samples The expected improvement value satisfies:

[0064] ;

[0065] in, The maximum objective function value in the current training sample set. The cumulative distribution function of the standard normal distribution. It is the probability density function;

[0066] when hour, Only hour It is positive, reflecting utilization;

[0067] when When it is large, This embodies exploration.

[0068] 2. Iterative optimization process: Initialization: Use the 100 samples output by the multi-scale defect model as the initial training set. Train the GPR model and calculate Iteration steps: with Given the objective function, the particle swarm optimization algorithm is used to search for the current optimal parameters. ;Will The multi-scale defect modeling module is input to generate corresponding dynamic simulation data and calculate the objective function value. ;Will Add to the training set and obtain Retrain the GPR model and update Repeat the steps until the preset number of iterations is reached.

[0069] 3.3 Generation and Output of Dynamic Simulation Data:

[0070] 1. Setting the impact load conditions: Simulating the impact load of aerospace collision avoidance scenarios, using a half-sine wave impact pulse, with a load peak value of... Pulse width The loading direction is perpendicular to the aluminum honeycomb core panel, and the load function satisfies:

[0071] .

[0072] 2. Contents of dynamic simulation data: Stress response data: Maximum principal stress of the macroscopic model. Minimum principal stress The sampling frequency is The time series length is Damage status data: Crack length in the microscopic model Failure rate of cohesive units Output format: Data will be output according to parameters. time The data is stored in a format that is CSV in type, and one CSV file is output for each iteration step.

[0073] Example: Parameter optimization of aerospace impact-resistant aluminum honeycomb core: Input conditions: Basic mechanical dataset output by the multi-scale defect modeling module, impact load parameters , Optimize target weights , Optimization process: Initialization: The kernel function parameters of the GPR model, containing 100 samples. , , 5th iteration: Found After inputting the multi-scale defect model, dynamic simulation data is obtained: The peak value is , The maximum value is , 20th iteration: Optimal parameters found The corresponding dynamic simulation data: , , Output: 20 iterations generated 20 CSV files, each containing 20 sets of parameter-dynamic simulation data; the optimal parameter vector is output. And the corresponding GPR proxy model.

[0074] Existing publicly available documentation on the application of Bayesian optimization in materials design only uses samples from ideal models to train surrogate models, without incorporating real data from multi-scale defect models. This leads to a high risk of optimization failure in practical applications. Existing publicly available documentation on aluminum honeycomb core parameter optimization only optimizes two parameters: cell wall thickness and pore size, without considering cell arrangement and without utilizing the EI function for balance exploration, resulting in low optimization efficiency. This technique offers the following advantages: Improved optimization efficiency: fewer iterations, shorter optimization cycle, and increased computational efficiency; Improved optimization effect: increased platform stress corresponding to optimal parameters, reduced damage rate, achieving a multi-objective balance of high load-bearing capacity and low damage; Increased data value: the generated dynamic simulation data contains the correlation between parameters, load, and damage, providing high-quality samples for subsequent LSTM model training, increasing the sample size.

[0075] IV. Construction of an LSTM-Driven Dynamic Damage Prediction Module: In existing technologies, dynamic damage prediction for aluminum honeycomb cores relies on empirical formulas, such as damage models derived from static compressed data or single sensor data. This fails to capture the nonlinear correlation between defect evolution, stress response, and damage accumulation under impact loads, resulting in prediction errors exceeding 18%. Therefore, it is necessary to construct an LSTM-based dynamic damage prediction module, using time-series data generated through Bayesian optimization as samples, to achieve real-time, high-precision damage prediction. Specific technical methods are as follows:

[0076] 4.1 Construction and labeling of the training dataset:

[0077] 1. Data Source and Preprocessing: Data Source: 20 sets of dynamic simulation data output by the Bayesian optimization module. Each set of data contains... Time series data; Data preprocessing: Missing value imputation: Linear interpolation is used to impute a small number of missing values ​​caused by simulation interruption; Normalization: Min-Max normalization is performed on the input features to satisfy: Time window partitioning: The time series is divided into samples using the sliding window method, with a window length of [missing information]. The step size is 10, and each set of dynamic data generates 996 samples. A total of 19,920 samples are generated from 20 sets of data.

[0078] 2. Label Definition and Symbols: Define the damage rate For tags, i.e. The damage state at any given time satisfies Define damage warning threshold ,when When this occurs, it is determined that the structure is about to fail, and a parameter correction signal needs to be triggered.

[0079] 4.2 LSTM Network Structure Design: Number of Layers and Node Settings: Input Layer: Input Feature Dimensions Window length Therefore, the shape of the input tensor is Hidden layer: Set up 2 layers of LSTM units, the number of neurons in the first hidden layer is... Number of neurons in the second hidden layer The activation function is tanh, and the dropout probability is set to 0.2; One fully connected layer is set, with a certain number of neurons. The activation function is ReLU; the output layer has a dimension of 1 and the activation function is sigmoid.

[0080] The forward propagation formula for networks:

[0081] Output of the first layer LSTM unit satisfy:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] Output of the second-layer LSTM unit Consistent with the first-layer structure, the input is ;

[0087] Output of fully connected layer and output layer satisfy:

[0088] ;

[0089] .

[0090] 4.3 Model Training and Drop Hammer Impact Test Calibration:

[0091] 1. Training parameter settings: Loss function: Use mean squared error, satisfying: Optimizer: Adam optimizer is used, learning rate Momentum parameters , Training rounds: After each training round, the model performance is evaluated using a validation set. Training is stopped when the validation set MSE does not decrease for 5 consecutive rounds.

[0092] 2. Calibration process for drop weight impact test: Test equipment: drop weight impact testing machine, hammer mass , fall high Impact speed Strain gauge arrangement: Six strain gauges are attached at different positions on the aluminum honeycomb core anti-collision panel. The sampling frequency of the strain gauges is [missing information]. The measurement direction is the Z-axis; damage state evaluation index: the strain value measured by the strain gauge. Converted into damage rate ,satisfy: Model calibration: This involves calibrating the experimental measurements. Predictions from the LSTM model By comparing the results, the MSE is calculated, and the weight parameters of the LSTM network are adjusted using the gradient descent method to reduce the MSE.

[0093] 4.4 Real-time Damage Prediction and Early Warning: Real-time Data Input: Real-time stress data from the macroscopic finite element model is input to the LSTM model via a data interface, with the input delay controlled within... Within; Predictive output and early warning logic: LSTM model each Output the predicted damage rate. and with threshold Compare; when When the signal is normal, output a normal signal and continue prediction; when At the same time, a damage warning signal is output, and the remaining bearing capacity is calculated: ; and will the warning signal with Push to the Bayesian optimization module.

[0094] Example: Dynamic damage prediction of aerospace collision-resistant aluminum honeycomb core: Input conditions: Optimal parameters output by the Bayesian optimization module. Strain data from drop hammer impact tests, real-time impact load Prediction process: Training phase: The LSTM model is trained with 19,920 samples, and the early stopping mechanism is implemented. Time-triggered, validation set MSE=0.0045; Calibration phase: converting the strain data from the drop hammer test into... Adjust the output layer weights of the LSTM , so that the predicted value and The deviation is reduced; in the real-time prediction stage: input corresponding , LSTM output ;when hour, LSTM output Trigger damage warning and calculate Output: Real-time output of damage rate time series. The warning signal is output in JSON format.

[0095] Existing publicly available documentation on machine learning applications in damage prediction only utilizes static image data and fails to incorporate dynamic simulation time-series data, thus failing to capture the damage evolution process. Existing publicly available documentation on LSTM applications in material property prediction does not employ drop-weight impact testing for model calibration, resulting in large prediction errors and the absence of real-time early warning logic. This technical approach offers the following advantages: Improved prediction accuracy: Reduced dynamic damage prediction errors and smaller prediction deviations in damage rates; Improved response speed: Shortened real-time prediction latency, enabling earlier triggering of damage warnings; Enhanced engineering applicability: Through drop-weight test calibration and residual load-bearing capacity calculation, the model can be directly used for health monitoring of aerospace collision avoidance structures, avoiding false or missed warnings.

[0096] V. Construction and Iterative Optimization of Closed-Loop Collaborative Architecture: In existing technologies, the three modules of multi-scale modeling, parameter optimization, and damage prediction operate independently, with data only transmitted unidirectionally (modeling → optimization) without feedback (optimization → modeling). This results in the model being unable to adjust parameters according to real-time damage status, leading to long design cycles and low reliability. Therefore, a closed-loop collaborative architecture of defect modeling, optimization, and damage prediction needs to be built to achieve bidirectional data interaction and iterative correction. Specific technical means are as follows:

[0097] 5.1 Design of Data Interaction Interface: Interface Type and Communication Protocol: Real-time communication between modules is achieved using the TCP / IP protocol, with port number set to 8080 and data transmission rate... The design includes three types of interfaces: Interface 1: transmits the initial parameter constraint range and basic mechanics dataset; Interface 2: transmits parameters—dynamic simulation data; Interface 3: transmits damage early warning signals. Timing control of data interaction is implemented: defining the timing period. Each cycle completes a closed loop of multi-scale modeling → Bayesian optimization → LSTM prediction → feedback correction; temporal logic: The multi-scale defect modeling module outputs basic data, which is then passed to the Bayesian optimization module via interface 1. The Bayesian optimization module completes 20 iterations and is then passed to the LSTM module via interface 2. The LSTM module completes training and prediction, and feeds back warning signals through interface 3; The Bayesian optimization module adjusts parameter constraints based on the warning signals, and the multi-scale modeling module updates the model to prepare for the next iteration.

[0098] 5.2 The core logic of closed-loop collaboration:

[0099] 1. Dynamic adjustment of parameter constraints: After receiving the warning signal from the LSTM, the Bayesian optimization module, if... Adjust the constraint range of the parameter to be optimized to the current parameter. To avoid parameter search entering high-damage regions; define the adjusted parameter constraint range: ; ;

[0100] 2. Real-time updating of the multi-scale model: The Bayesian optimization module will adjust the parameters... The input is passed to the multi-scale defect modeling module, which automatically updates the cell wall thickness and pore size of the microscopic cell model and regenerates the macroscopic model; update logic: if Compared with the previous round of parameters deviation If the deviation is found, only the material parameters will be updated; if the deviation is found, the material parameters will be updated. Then the Voronoi micromodel is regenerated.

[0101] 5.3 Setting the iteration termination condition:

[0102] 1. Quantitative definition of termination condition: Condition 1: Prediction error of mechanical properties in three consecutive iterations ,in: Condition 2: The damage warning signal output by the LSTM module is stable for three consecutive rounds; Condition 3: Number of iterations. .

[0103] 2. Iteration termination determination process: After each iteration, calculate... , deviation, Deviation; if both conditions 1 and 2 are met, or condition 3 is met, stop the iteration and output the final model; otherwise, proceed to the next iteration and repeat the process of data interaction → parameter optimization → model update.

[0104] Example: Closed-loop iterative optimization of aerospace collision-resistant aluminum honeycomb cores: Input conditions: initial multi-scale defect model, initial parameter range of Bayesian optimization, warning threshold of LSTM. Iteration termination condition , Iterative process: First iteration: Multi-scale modeling outputs basic data, which is then optimized using Bayesian methods. LSTM prediction , Third iteration: Bayesian optimization adjusts parameter constraints based on the LSTM's warning signal, resulting in... LSTM prediction , Fourth iteration: Multi-scale modeling updates the model, and Bayesian optimization is performed. LSTM prediction , If condition 2 is also met, stop the iteration. Output: Final finite element model, model parameter report, output file size approximately 1GB.

[0105] Existing publicly available documents describe multi-module collaborative design, which only implements a one-way process of modeling → optimization → analysis, lacking closed-loop feedback and unable to adjust parameters based on real-time damage. Furthermore, existing publicly available documents describe aluminum honeycomb core design methods without setting quantitative iteration termination conditions, relying on manual judgment, resulting in long design cycles. This technical solution offers the following advantages: shortened design cycle; improved model reliability; enhanced engineering application value: the closed-loop architecture can be directly integrated into the digital design platforms of aerospace enterprises, enabling one-click design → simulation → early warning of aluminum honeycomb core collision protection structures, thus improving the design pass rate.

[0106] In summary, this technical solution, using aluminum honeycomb core collision-resistant structures in the aerospace field as an example, achieves high-precision and high-efficiency design of finite element models of the mechanical properties of aluminum honeycomb cores through a complete process: multi-scale defect coupling modeling → Bayesian optimization for dynamic parameter correction → LSTM-driven dynamic damage prediction → closed-loop collaborative iteration. Specifically, multi-scale defect modeling addresses the issue of physical realism, Bayesian optimization solves the problem of parameter optimization efficiency, LSTM prediction addresses the problem of dynamic damage early warning, and the closed-loop architecture solves the problem of module collaboration. The combined technical solution, compared with existing publicly available documents, significantly improves modeling accuracy, design efficiency, and engineering practicality, and can be widely applied to the design of aluminum honeycomb core structures in aerospace, new energy vehicles, and rail transportation fields.

Claims

1. A design method for a finite element model of the mechanical properties of an aluminum honeycomb core, characterized in that: Includes the following steps: Step 1: Construct a multi-scale defect coupling modeling module to extract manufacturing defect features of aluminum honeycomb cores and generate a multi-scale geometric model containing defects, and output the initial geometric model and basic mechanical dataset; Step 2: Based on the basic mechanics dataset, construct a Bayesian optimization dynamic parameter correction module, and generate dynamic simulation data by iteratively searching for the optimal parameter combination through a surrogate model. Step 3: Using dynamic simulation data as input, construct an LSTM-driven dynamic damage prediction module, train the model, and realize real-time damage state prediction. Step 4: Establish a closed-loop collaborative architecture of the multi-scale defect coupling modeling module, the Bayesian optimization dynamic parameter correction module, and the LSTM-driven dynamic damage prediction module through the data interaction interface. The damage warning signal output by the LSTM-driven dynamic damage prediction module serves as the dynamic constraint condition of the Bayesian optimization dynamic parameter correction module. The parameter combination output by the Bayesian optimization dynamic parameter correction module drives the multi-scale defect coupling modeling module to update the model. After iterating until the preset conditions are met, the final finite element model of the mechanical properties of the aluminum honeycomb core is output. In step four, the data interaction interface establishment process includes: encapsulating the defect rate and cell wall thickness deviation parameters output by the multi-scale defect coupling modeling module into the initial parameter constraint range of the Bayesian optimization dynamic parameter correction module; pushing the parameter-mechanical response-damage state data generated by the Bayesian optimization dynamic parameter correction module to the training dataset of the LSTM-driven dynamic damage prediction module; when the damage rate predicted by the LSTM-driven dynamic damage prediction module exceeds the preset threshold, generating an emergency parameter correction signal, which is transformed into the dynamic constraint condition of the Bayesian optimization dynamic parameter correction module. The specific transformation rule is: adjusting the value range of the parameter to be optimized to the current optimal parameter value ± the preset adjustment step size, wherein the preset adjustment step size is set according to the parameter type, the cell wall thickness adjustment step size is 5% of the nominal cell wall thickness, and the pore size distribution adjustment step size is 3% of the nominal pore size.

2. The design method of the finite element model of the mechanical properties of aluminum honeycomb core as described in claim 1, characterized in that: In step one, the construction process of the multi-scale defect coupling modeling module includes: using microscopic imaging technology to collect manufacturing defect features of aluminum honeycomb cores, including cell wall defects, burrs, and bonding defects; generating a micro-cell model containing manufacturing defect features based on the improved Voronoi algorithm, and introducing a cohesive force model to define crack propagation criteria at defects; mapping the defect features of the micro-cell model to the macroscopic finite element model through sub-model technology to obtain the initial geometric model; performing static compression simulation on the initial geometric model and outputting a basic mechanics dataset, which includes static compressive stress-strain curves under defect-free and defect-containing states.

3. The design method of the finite element model of the mechanical properties of aluminum honeycomb core as described in claim 2, characterized in that: In step two, the construction process of the Bayesian optimization dynamic parameter correction module includes: using the basic mechanics dataset as a sample, constructing a Gaussian process regression surrogate model; defining the optimization objective as maximizing platform stress and minimizing damage rate, determining the range of parameters to be optimized, including cell wall thickness and pore size distribution; using the expectation improvement acquisition function to iteratively search for the optimal combination of parameters to be optimized, and in each iteration, inputting the output combination of parameters to be optimized into the multi-scale defect coupling modeling module to generate dynamic simulation data under the corresponding parameters, including stress response data under impact load.

4. The design method of the finite element model of the mechanical properties of aluminum honeycomb core as described in claim 1, characterized in that: In step three, the construction process of the LSTM-driven dynamic damage prediction module includes: labeling the dynamic simulation data in a time-series format to form a training dataset for the LSTM model, which includes defect parameters, load history, and damage evolution data; training the LSTM network based on the training dataset to enable the LSTM network to capture the nonlinear correlation between defect parameters, load history, and damage evolution; calibrating the trained LSTM network using strain gauge data obtained from drop hammer impact tests to ensure that the damage prediction error of the LSTM network meets preset requirements; and embedding the calibrated LSTM network into the finite element simulation process to output the damage state prediction results in real time, which include crack propagation rate and remaining bearing capacity.

5. The design method of the finite element model of the mechanical properties of aluminum honeycomb core as described in claim 1, characterized in that: In step four, iterating until the preset conditions are met specifically involves: setting the iteration period and the prediction error threshold, and recording the mechanical performance prediction error of the final finite element model in each iteration. When the mechanical property prediction error of three consecutive iterations is less than the prediction error threshold, and the damage warning signal output by the LSTM-driven dynamic damage prediction module is stable, the iteration stops and the final finite element model of the mechanical properties of the aluminum honeycomb core is output.

6. The design method of the finite element model of the mechanical properties of aluminum honeycomb core as described in claim 2, characterized in that: The improvement of the improved Voronoi algorithm lies in the introduction of actual manufacturing process parameters of aluminum honeycomb core to constrain the generation boundary of the Voronoi diagram. The actual manufacturing process parameters include cell wall forming error and cutting accuracy parameters.

7. The design method of the finite element model of the mechanical properties of aluminum honeycomb core as described in claim 3, characterized in that: The parameters to be optimized also include the cell arrangement of the aluminum honeycomb core. During the iterative search process, the Bayesian optimization dynamic parameter correction module synchronously records the differences in dynamic simulation data under different cell arrangements and feeds the difference data back to the multi-scale defect coupling modeling module to optimize the arrangement structure of the micro-cell model.

8. The design method of the finite element model of the mechanical properties of aluminum honeycomb core as described in claim 4, characterized in that: When calibrating an LSTM network using a drop hammer impact test, strain data is collected by attaching strain gauges at different locations on an aluminum honeycomb core. The strain data is then converted into damage state evaluation indicators, which are compared with the prediction results of the LSTM network. The weight parameters of the LSTM network are then adjusted to reduce prediction errors.

Images