CN-121999931-A - Design method for brittle material with toughening and controllable crack path

CN121999931ACN 121999931 ACN121999931 ACN 121999931ACN-121999931-A

Abstract

The invention discloses a design method for integrating toughness improvement and crack path controllability aiming at brittle materials, which belongs to the technical field of fracture mechanics and artificial intelligence intersection and comprises the steps of carrying out feature extraction by adopting an input coding method based on space-response double-domain fusion, carrying out weighted fusion on extracted double-domain features by introducing a channel attention mechanism, generating prediction results aiming at different key tasks, evaluating the reliability of the prediction results by adopting a phase field verification and closed loop self-adaptive optimization mode, carrying out iterative training and optimization on a neural network, and carrying out multi-objective optimization on defect generation parameters by taking the optimized neural network as an adaptability evaluation function of double-objective genetic optimization. The method and the device have the advantages that geometric forms and dynamic response characteristics are cooperatively extracted through fusion of the space domain and the response domain, prediction accuracy of complex fracture behaviors is improved, and optimal solutions with high toughness and controllable paths are output through non-dominant sequencing multi-objective optimization under physical constraint, so that brittle materials are pushed to be spanned to active controllable fracture designs.

Inventors

CAO XIAOFEI
Mei Ruoyan
ZHU SIRONG
KANG XIAO
HE CHUNWANG
WAN ZHISHUAI
CAO MIAO
WANG ZHUANGZHUANG

Assignees

武汉理工大学
北京理工大学

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (10)

1. A design method for integrating toughening and crack path control for brittle materials is characterized by comprising the following steps: s1, performing feature extraction by adopting an input coding method based on space-response double-domain fusion; S2, introducing a channel attention mechanism to perform weighted fusion on the double-domain features extracted in the S1, and generating prediction results for different key tasks; s3, evaluating the reliability of the prediction result by adopting a phase field verification and closed loop self-adaptive optimization mode, and performing iterative training and optimization on the neural network; and S4, taking the neural network subjected to the optimization of the S3 as an adaptability evaluation function of the double-target genetic optimization, and performing multi-target optimization on the defect generation parameters.
2. The design method for achieving both toughening and crack path controllability for brittle materials according to claim 1, wherein the following are specifically included in S1: S11, optimizing variable definition and target path presetting, namely generating a batch of individuals by a genetic algorithm, wherein each individual is a sine groove, and a central line equation is as follows: ; Wherein, the In order for the amplitude to be the same, As a function of the wavelength(s), Is phase offset, the design parameter is amplitude for controlling the fluctuation degree of the groove Wavelength of groove period length is controlled Width of groove Groove thickness Thereby defining the optimized parameter set as Then presetting a target path, and converting the target path into a target path matrix P target for the calculation of a subsequent second target function; S12, data input of a mathematical definition space domain and a response domain specifically comprises the following steps: S121, in the plane area Defining a binary defect distribution function: ; Wherein, the Is the center line of the defect area when The time is shown in With defects in position, when The time is shown in No defect in position; S122, function is added Discrete as a rasterized matrix The method is used for 2D-CNN input of a two-dimensional convolutional neural network and represents defect distribution of a sample in a spatial domain, wherein, Representing the height of the discretized matrix, Representing the width of the matrix after discretization, Each element of (a) is the corresponding pixel position Finally, a two-dimensional matrix with only 0 and 1 elements is obtained, wherein 0 represents normal, and 1 represents defect; s123, extracting features on a defect distribution map by using a convolutional neural network CNN, wherein the features comprise position modes and shapes of defects for further learning, and receiving a discretized defect distribution matrix by an input layer Each layer of convolution is followed by a maximum pooling layer to reduce the dimension and extract key geometric features; s124, at time step Under the control of displacement, loading and calculating counter force Obtaining force-displacement curves, i.e. Wherein In order for the displacement to be a function of, Is the corresponding counter force; S125, taking the counterforce sequence and normalizing to be Obtaining a counterforce sequence with the length of T after normalization The method is used for inputting 1D-CNN of the one-dimensional convolutional neural network; S126, 1D-CNN adopts a Leaky ReLU activation function to maintain the nonlinear mapping capability of the model; s13, extracting spatial domain and response domain characteristics, namely respectively adopting 2D-CNN and 1D-CNN to correspondingly process spatial domain and response domain input, wherein the spatial domain and response domain characteristics are as follows: for spatial domain feature extraction, a two-dimensional convolution kernel set is first defined And is input in space Performing two-dimensional convolution operation on a convolution object, wherein the calculation process is as follows Wherein the symbols are Representing two-dimensional convolution operations As a function of the non-linear activation, Is the first The bias terms of the individual convolution kernels, Represent the first Spatial domain features obtained by the layers; For response domain feature extraction, a one-dimensional convolution kernel set is defined Inputting normalized response vector One-dimensional convolution operation is carried out, and the expression is as follows Wherein the symbols are A one-dimensional convolution operation is represented, In order to activate the function, Is the first The bias terms of the individual convolution kernels, Then represent the first Response domain characteristics of the layer.
3. The design method for achieving both toughening and crack path controllability for brittle materials according to claim 2, wherein in S2, the following are specifically included: S21, introducing channel attention mechanism The double-domain features are weighted and fused, specifically: s211, spatial feature is to be generated And response characteristics Converted into one-dimensional vectors by flattening operations, i.e , ; S212, splicing to form a joint feature vector ; S213, constructing an attention weight module, and introducing a weight coefficient And Wherein And As a weight vector that can be trained, Respectively representing the weight proportion of the space domain and the response domain in the integral characteristic representation, and meeting ; S214, acquiring comprehensive representation after fusing the spatial structure and loading response characteristics according to the attention weighting result , As input to the subsequent fully connected layer and prediction module; s22, fusion features Through full-connection mapping, prediction results are generated for different key tasks respectively, so that comprehensive characterization of crack evolution behaviors is realized, and the method specifically comprises the following steps: S221, outputting a crack path probability map by using the full-connection layer with the Sigmoid activation function aiming at the spatial distribution of the crack path Wherein each element value Representing the probability that the corresponding position becomes a crack path, and visually presenting the space trend of the crack; S222, through linear mapping Obtaining fracture toughness indexes, and evaluating the overall fracture performance of the material structure; S223, setting key parameters of a branch prediction model output by utilizing Softplus activation functions, namely characteristic length, aiming at parameter solving requirements of a phase field fracture model In the form of Thereby directly characterizing the physical mechanism of the material fracture process.
4. The design method for achieving both toughening and crack path controllability for brittle materials according to claim 3, wherein in S3, the following are specifically included: S31, introducing a random node shielding mechanism to perform multi-task uncertain output; s32, establishing a sample screening mechanism based on the fluctuation of the prediction result to screen samples, and screening out samples with high uncertainty; S33, performing phase field high-precision verification on the high uncertainty sample in the S32; S34, dynamically adjusting the weight of the spatial domain features and the response domain features in fusion according to the error type of the high uncertainty sample, and then retraining the network by utilizing the adjusted features; s35, setting iteration termination conditions to obtain a corrected neural network model.
5. The design method for achieving both toughening and crack path controllability for brittle materials according to claim 4, wherein the design method comprises the following steps: in S31, the specific contents are: S311, carrying out the same input sample And performing forward computation, wherein partial nodes are continuously and randomly shielded in each computation, so that the neuron combination participating in the operation is continuously changed, and the current input corresponding is obtained Predicted values for structural differences of the groups; S312, according to the formula And Respectively calculating the average value and variance of the predicted value of S311, wherein the average value is taken as expected output, and the variance is a quantization index for the prediction stability; s32, specifically, setting a threshold value for controlling fluctuation amplitude When the fluctuation amplitude of the predicted sample satisfies When the output is determined to be stable and reliable, the average value is adopted The crack path and the fracture performance are solved by taking the crack path and the fracture performance as input parameters of a phase field analysis module, when the fluctuation range meets the requirement When a significant instability of the results is demonstrated, such samples are classified as high uncertainty samples; In S33, the specific contents are: s331, carrying out defect geometric parameters of high uncertainty samples Outputting the predicted phase field characteristic parameters Input phase field fracture finite element module Solving the total free energy functional of the material in the module , Obtaining a true path of the crack in the solid domain Fracture toughness ; S332, crack path diagram to be predicted True path diagram obtained by phase field method Adopts the formula of the cross-over ratio A two-dimensional image comparison is performed and, Measuring the consistency of crack positions and expansion forms; S333, obtaining predicted fracture toughness Calculated value of phase field Is the relative difference of (2) Namely physical consistency errors, are used for evaluating the deviation of the material global mechanical property prediction; S334, comparing the characteristic length parameters of the prediction of the adjacent two rounds Monitoring the stability and convergence trend of the model in continuous iteration; S335, error of path Error in performance And uniformly writing the parameter variation into the self-adaptive weight correction module for closed-loop feedback, and adjusting the network weight and the self-adaptive parameters to ensure that the model continuously converges in iteration, synchronously improving the prediction precision and keeping the physical consistency.
6. The method for designing a brittle material with both toughness and crack path controllability according to claim 5, wherein S34 comprises the following steps: S341, updating the branch weight, namely, the path error read from the last module And performance error Back press And The spatial domain and response domain weights are adjusted, respectively, wherein, For the spatial domain branch weights, In response to the domain branch weights, And The weight is increased when the error is larger, the prompt model pays attention to the field with serious error, and the network pays attention to the corresponding characteristic; s342, updating fusion layer expression, and substituting new weight into formula Recombining the feature fusion layers to form updated feature vectors ; S343, retraining the network by utilizing the new fusion characteristics and the updated sample set Retraining the neural network to obtain a model 。
7. The method for designing a brittle material with both toughness and crack path controllability as claimed in claim 6, wherein S35 comprises the following specific contents in the first step After the round iteration is finished, recording the path error Error in performance Parameter variation Setting a path accuracy threshold Performance accuracy threshold And parameter variation tolerance Three precision tolerances, when simultaneously satisfying , , And Stopping the circulation, feeding back the corresponding error signal to the weight correction module if any condition is not met, continuing the next iteration, regarding that the model is converged when the termination condition is met, and outputting the final crack path prediction, fracture performance and phase field parameter results, wherein the network prediction and physical solution errors are in a preset tolerance range at the moment, and the closed loop iteration process is finished.
8. The method for designing a brittle material with both toughness and crack path controllability according to claim 7, wherein S4 comprises the following steps: S41, constructing a comprehensive evaluation system comprising fracture toughness maximization, path deviation minimization and physical consistency constraint by using a trained two-domain fusion neural network as a rapid evaluation proxy model; S42, constructing a comprehensive fitness function comprising a normalization weighting and penalty function mechanism ; S43, randomly generating a plurality of groups in the parameter space As an initialization population, each geometrical parameter set is encoded into a chromosome, and after selection, crossover, variation probability and iteration times are set, a virtual response sequence is rapidly generated by geometrical parameters by utilizing a pre-trained lightweight mapping network to complement the input of a double domain; s44, non-dominant sorting and elite reservation are carried out; S45, stopping iteration when the target difference between the current generation and the previous generation meets a set threshold and reaches the maximum iteration number, and considering that the algorithm reaches stable convergence; S46, outputting a parameter set on the Pareto front when the population evolves to meet the convergence condition I.e. optimal solutions which cannot simultaneously further improve toughness without increasing path error, wherein each optimal solution corresponds to a pair of performance indexes Optimum result Directly used as the geometric design input parameter of the material defect.
9. The design method for achieving both toughening and crack path controllability for the brittle material according to claim 8, wherein in S41, the specific construction is as follows: S411, setting a first objective function to maximize predicted fracture toughness ; Scalar results output by the neural network performance prediction branches are directly extracted as fracture toughness indexes of the materials, Setting an optimization target as a maximum fracture toughness index for a fracture toughness prediction result directly output by the neural network model according to input characteristics: , wherein, A defect geometric parameter set corresponding to the current genetic algorithm individual, Representing the performance prediction mapping relation of the neural network; S412, setting a second objective function, namely minimizing the path probability distribution deviation ; Crack path probability distribution matrix output by neural network And a preset target path matrix Performing pixel level difference calculation, and setting an optimization target to minimize distribution difference between the two The objective function drives the optimization process to screen out a design scheme with induced cracks extending along a preset track accurately and with clear probability distribution; s413, setting physical consistency constraint, namely penalty term; introducing characteristic length parameters based on neural network output Removing false solutions meeting numerical targets but violating physical laws, setting effective physical parameter intervals If the characteristic length of the network prediction exceeds the interval, the network prediction is regarded as physical failure, and a penalty function is defined as When the physical parameters predicted by the neural network exceed the effective physical interval, the fitness of the individual is forcedly reduced, so that the individual is naturally eliminated in the evolution process.
10. The method for designing a brittle material with both toughness and crack path controllability as claimed in claim 9, wherein S42 comprises the following steps of Deviation from the path Normalized, and the comprehensive fitness function is defined as Wherein And Is a weight coefficient, satisfies Design preferences for adjusting the propensity for high toughness and propensity for path accuracy, And As the estimated maximum value for normalization, The phase field characteristic width is output for the neural network; s44, specifically comprising the following steps: S441, combining a parent population with a child population generated by cross mutation to form a mixed population, assigning an individual non-dominant grade according to an fitness function before sorting, and directly assigning the lowest non-dominant grade if the fitness of the individual is classified as a punishment value so as to naturally eliminate the individual in subsequent selection; s442, quick non-dominant sorting according to objective function And Layering the effective individuals in the mixed population, defining individual i to dominate individual j, and recording as If and only if it meets Dividing the population into a plurality of non-dominant layers according to the dominant relationship, namely, the level 1 and the level 2, wherein individuals in the level 1 are mutually not dominant to form the current Pareto front; S443, calculating the crowding degree distance, namely calculating the crowding degree distance of the individuals in the same non-dominant level, carrying out normalized sequencing on the numerical values of each objective function, setting the crowding degree of the boundary individuals as infinity, wherein the crowding degree of the middle individuals is the sum of the distances of two adjacent individuals in the target direction; S444, elite reservation and new generation population generation, namely constructing a next generation population based on the principle of level priority and distance order, firstly, sequentially placing individuals with high level levels such as level 1, level 2 and the like into the new population, when the population quantity exceeds a preset scale N due to the fact that the individuals are placed in a certain level, cutting off and selecting the individuals in the level according to the crowding degree distance from large to small, and finally obtaining N reserved individuals as the new generation population.

Description

Design method for brittle material with toughening and controllable crack path Technical Field The invention relates to the technical field of fracture mechanics and artificial intelligence intersection, in particular to a design method for considering both toughening and crack path controllability of brittle materials. Background Brittle materials are extremely sensitive to crack initiation and propagation due to internal defects, often exhibit high modulus of elasticity and lower fracture toughness, and tend to fracture instantaneously without signs after being stressed. To ameliorate this weakness, engineering has often employed a "structure-guided" strategy, i.e., to actively design specific microstructures, such as grooves, within the material, to induce crack propagation along a predetermined path, thereby consuming more energy and avoiding catastrophic failure. Thus, how to predict, control and design crack paths becomes critical for toughening. In the aspect of crack path prediction, the prior art mainly relies on two types of methods, namely a phase field method and a deep learning method, but the two types of methods still have a plurality of problems. The existing deep learning method is generally limited by single-dimensional input, and the static defect form and the dynamic mechanical response are forcedly processed separately, so that the model is difficult to capture the complex coupling effect of the static defect form and the dynamic mechanical response in the fracture evolution process. Meanwhile, the conventional model can only give unidirectional prediction and cannot self-evaluate uncertainty due to the lack of a statistical mechanism for quantifying the reliability of the result, so that the system is difficult to perform effective risk early warning and self-adaptive optimization when facing unknown working conditions. More importantly, if the constraint of a physical mechanism is lost, the prediction result of the pure data driving network is easy to violate the actual fracture rule, so that the precision is greatly reduced when complex behaviors such as crack competition, bifurcation and the like are processed. The phase field method has the advantages that a complex evolution process from crack initiation to bifurcation can be automatically simulated without presetting a path, and because the phase field method can provide a high-precision solution which strictly meets the physical conservation law, a key mechanism constraint and an error correction basis are provided for a data driving model. However, as the microstructure design becomes more complex, this conventional method reveals its limitations in that not only is the calculation cost extremely high, it is difficult to cope with the screening of a large number of design parameters, but also the simulation accuracy thereof is highly dependent on the setting of the characteristic width parameters. In the prior art, the parameter is fixed by adopting an empirical method or a trial and error method, and when the microstructure or defect morphology is changed, the optimal characteristic width is also required to be adjusted accordingly. The setting mode not only increases the burden of manual repeated debugging, but also cannot ensure that parameters are completely matched with the model, thereby influencing the calculation accuracy. In terms of structural optimization design, most designs only take the improvement of fracture toughness as a unique optimization index, and attempt to delay the occurrence of fracture by changing structural parameters. However, such single-target-oriented designs tend to ignore the critical issue of crack propagation path controllability. Even for multi-objective optimization designs, the prior art often uses a simple linear weighting method to force the fracture toughness and crack path deviation together into a scalar objective. This approach has significant drawbacks. First, the setting of the weight coefficients is very subjective, and it is difficult for the designer to predict the mathematical relationship between "toughness" and "path" before optimization, so that the optimization result tends to deviate from one party to the other party and sacrifice the other party, and a true equilibrium solution cannot be obtained. Second, conventional genetic algorithms lack real-time constraint mechanisms of physical mechanisms. In a purely data driven iterative process, the algorithm is extremely prone to generating an ineffective design that geometrically satisfies mathematical extrema, but physically violates the material fracture characteristic length. These "pseudo-optimal solutions" often need to be culled later through expensive finite element calculations, greatly wasting computational resources and reducing the convergence efficiency of the design. Disclosure of Invention The invention aims to provide a design method for considering both toughening and crack path controllability of brittle mater