CN-122024942-A - Multi-target particle swarm optimization crystal plastic model parameter calibration method based on coupling machine learning

CN122024942ACN 122024942 ACN122024942 ACN 122024942ACN-122024942-A

Abstract

The invention relates to a multi-target particle swarm optimization crystal plastic model parameter calibration method based on coupling machine learning, which comprises the steps of S1, generating a parameter sample set through Latin hypercube sampling based on a parameter range of a crystal plastic model, and obtaining a stress strain curve by utilizing crystal plastic finite element simulation, S2, training through an artificial neural network ANN based on the parameter sample set of S1 and a corresponding stress strain curve serving as a machine learning data set, and constructing a data-driven proxy model, and S3, coupling the proxy model of S2 with a multi-target particle swarm optimization algorithm, and obtaining an optimal parameter combination meeting the multi-orientation experimental data consistency through iterative optimization. According to the method, the data are generated through finite elements, the agent model is replaced, the multi-objective optimization iteration is carried out, the calculation cost is reduced, the artificial neural network is trained for different material orientations, and the prediction pertinence is improved.

Inventors

GAO HONG
HU JIAQI
WU YUCHEN
Sun xingyue

Assignees

天津大学

Dates

Publication Date: 20260512
Application Date: 20251219
Priority Date: 20251210

Claims (8)

1. A method for calibrating parameters of a crystal plastic model based on multi-target particle swarm optimization of coupling machine learning is characterized by comprising the following steps: S1, generating a parameter sample set through Latin hypercube sampling based on a parameter range of a crystal plastic model, and obtaining a stress-strain curve by utilizing crystal plastic finite element simulation; s2, training by using the initial parameter sample set generated based on the S1 and a corresponding stress-strain curve as a machine learning data set through an artificial neural network ANN, and constructing a data-driven proxy model; And S3, coupling the agent model of the S2 with a multi-target particle swarm optimization algorithm, and obtaining an optimal parameter combination meeting the consistency of the multi-orientation experimental data through iterative optimization.
2. The method for calibrating parameters of a crystal plastic model based on multi-objective particle swarm optimization of coupled machine learning according to claim 1, wherein the parameter range of S1 considers three sliding systems of basal < a > sliding ({ 0001} <11-20 >), cylindrical < a > sliding ({ 10-10} <11-20 >) and II-order conical surface < c+a > sliding ({ -1-122} < -1-123 >) and a stretching twin ({ 10-12} < -1011 >) system; Latin hypercube sampling fills 20 parameter spaces, 100 groups of parameter sample sets are generated by sampling, the parameter sample sets are used for carrying out crystal plasticity finite element simulation calculation on the samples, 100 groups of stress-strain curves are obtained, non-converged simulation data are removed, and smooth pretreatment is carried out on the stress-strain curves obtained through simulation.
3. The method for calibrating the parameters of the crystal plastic model based on the multi-objective particle swarm optimization of the coupling machine learning according to claim 1, wherein the artificial neural network of S2 adopts a single hidden layer structure, the number of neurons of the hidden layer is 128, the activation function is LeakyReLU, the optimization algorithm is Adam, and the loss function is mean square error.
4. The method for calibrating parameters of a coupled machine learning based multi-objective particle swarm optimized crystal plastic model according to claim 1, wherein S2 uses additional samples to verify the proxy model, and stress-strain responses of each sample orientation are predicted by an independent ANN model.
5. The method for calibrating the parameters of the crystal plastic model based on the multi-objective particle swarm optimization of the coupling machine learning according to claim 1, wherein the multi-objective particle swarm optimization algorithm of the S3 takes an R 2 value as an fitness function to evaluate the consistency of a simulated and experimental stress-strain curve, the initialized particle swarm of the multi-objective particle swarm optimization algorithm fills initial particle swarm coordinates P 0 in an LHS mode in a specified crystal plastic parameter range, the number of particles is 100, an initial speed V0 is randomly given to the population, the reference speed range is set to be +/-0.3 times of the parameter boundary width, and the S3 uses a trained ANN proxy model to replace the crystal plastic finite element model, inputs the particle swarm coordinates P 0 and outputs the stress-strain curve.
6. The method for calibrating parameters of a crystal plastic model based on coupled machine learning multi-objective particle swarm optimization according to claim 5, wherein S3 uses R 2 to evaluate the consistency (fitness) between the actual single pull orientation of each sample and the corresponding simulation result, and the expression is: ; (1) Wherein: And Representing simulated and experimental stresses under plastic deformation, Is the average value of the simulated stress, and according to the adaptability result, an external archive is created to store all the current non-dominant solutions, namely the Pareto (Pareto) front, and the solution diversity is ensured by adopting the adaptive grid screening through the Pareto front storage non-dominant solutions.
7. The method for calibrating parameters of a crystal plastic model based on coupled machine learning multi-objective particle swarm optimization according to claim 5, wherein S3 updates the particle velocity and position according to the optimal position (particle optimal solution) of the previous t steps for the ith particle And global optimum position of previous t steps The (global optimal solution) updates its velocity and position, the global optimal solution is selected from the non-dominant solutions of the pareto fronts, the update equation is as follows: ; (2) ; (3) Wherein: And The coordinate and the velocity vector corresponding to the ith particle in the t step are respectively, and omega, c 1 and c 2 are super parameters.
8. The method for calibrating parameters of a crystal plastic model based on multi-objective particle swarm optimization of coupled machine learning according to claim 1, wherein the method is applicable to the calibration of parameters of materials, magnesium alloys or titanium alloys with crystal plastic deformation behavior.

Description

Multi-target particle swarm optimization crystal plastic model parameter calibration method based on coupling machine learning Technical Field The invention belongs to the technical field of simulation of mechanical properties of materials, and particularly relates to a multi-target particle swarm optimization crystal plastic model parameter calibration method based on coupling machine learning. Background The Crystal Plasticity Finite Element Model (CPFEM) is an important tool for researching the mechanical behavior of the polycrystalline material, but the parameter calibration depends on a large number of trial and error methods or traditional optimization algorithms, and has the problems of high calculation cost and great manual intervention. The traditional method relies on manual experience to repeatedly adjust parameters, multiple times of running finite element simulation and comparing experimental data are needed, time consumption and subjectivity are high, the traditional optimization algorithm such as a gradient descent method, a genetic algorithm and the like needs to frequently call the finite element simulation, single simulation takes more than 5 minutes, the optimization process needs tens of thousands of iterations, the total time consumption can reach tens of thousands of minutes, the calculation cost is too high, samples with different material orientations need to be simultaneously optimized for a plurality of conflict targets (such as a plurality of characteristics of stress-strain curves), the traditional single-target optimization algorithm cannot be effectively used for processing, and the multi-target particle swarm algorithm (MOPSO) can generate the pareto front, but the method relies on high-frequency simulation calculation, so that the calculation resource requirement is increased explosively. In order to solve the problems, the invention provides a staged parameter calibration method combining a finite element simulation, a machine learning agent model and a multi-objective optimization algorithm. Disclosure of Invention Aiming at the defects and shortcomings of the prior art, the invention provides a multi-target particle swarm optimization crystal plastic model parameter calibration method based on coupling machine learning, which remarkably improves efficiency and precision. The invention solves the technical problems by the following technical proposal: a method for calibrating parameters of a crystal plastic model based on multi-target particle swarm optimization of coupled machine learning comprises the following steps: S1, generating a parameter sample set through Latin hypercube sampling based on a parameter range of a crystal plastic model, and obtaining a stress-strain curve by utilizing crystal plastic finite element simulation; s2, training by using the initial parameter sample set generated based on the S1 and a corresponding stress-strain curve as a machine learning data set through an artificial neural network ANN, and constructing a data-driven proxy model; And S3, coupling the agent model of the S2 with a multi-target particle swarm optimization algorithm, and obtaining an optimal parameter combination meeting the consistency of the multi-orientation experimental data through iterative optimization. Moreover, the parameter range of S1 considers three sliding systems of basal plane < a > sliding ({ 0001} <11-20 >), cylindrical plane < a > sliding ({ 10-10} <11-20 >) and II order conical surface < c+a > sliding ({ -1-122} < -1-123 >) and one stretching twin ({ 10-12} < -1011 >) system; Latin hypercube sampling fills 20 parameter spaces, 100 groups of parameter sample sets are generated by sampling, the parameter sample sets are used for carrying out crystal plasticity finite element simulation calculation on the samples, 100 groups of stress-strain curves are obtained, non-converged simulation data are removed, and smooth pretreatment is carried out on the stress-strain curves obtained through simulation. Moreover, the artificial neural network of the S2 adopts a single hidden layer structure, the number of neurons of the hidden layer is 128, the activation function is LeakyReLU, the optimization algorithm is Adam, and the loss function is mean square error. Moreover, the S2 uses additional samples to validate the proxy model, with stress-strain responses for each sample orientation predicted by a separate ANN model. The multi-target particle swarm optimization algorithm of the S3 takes an R 2 value as an fitness function to evaluate consistency of a simulated and experimental stress-strain curve, an initial particle swarm coordinate P 0 is filled in an LHS mode in a specified crystal plastic parameter range by the initialized particle swarm of the multi-target particle swarm optimization algorithm, initial speed V0 is randomly given to the particle swarm, a reference speed range is set to be +/-0.3 times of a parameter boundary width, and the S3 uses a trained ANN