CN-122021125-A - Metal stress strain constitutive relation prediction method and system based on machine learning

CN122021125ACN 122021125 ACN122021125 ACN 122021125ACN-122021125-A

Abstract

The invention belongs to the technical field of material mechanical property modeling and machine learning intersection, and particularly relates to a metal stress strain constitutive relation prediction method and system based on machine learning, wherein the metal stress strain constitutive relation prediction method based on machine learning comprises the following steps of obtaining a training sample set, training a machine learning model by using the training sample set, and optimizing super parameters of the machine learning model by combining a Bayesian optimization algorithm to obtain a prediction model; and acquiring the working condition temperature, the working condition strain and the working condition strain rate of the target metal under the working condition to be detected, and inputting the working condition temperature, the working condition strain and the working condition strain rate into the prediction model to obtain the stress predicted value and the stress strain constitutive relation of the target metal under the working condition to be detected. The prediction method provided by the invention predicts the precision And the experimental cost is reduced by 80 percent, and the method is suitable for hot forming and mechanical simulation scenes of aerospace structural parts.

Inventors

YE DONG
ZHANG AIBIN
ZHANG CHEN
QIE XIWANG
ZHANG MEIJUAN

Assignees

北京航空材料研究院股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. The metal stress strain constitutive relation prediction method based on machine learning is characterized by comprising the following steps of: acquiring a training sample set, wherein each training sample in the training sample set independently comprises input parameters and output parameter data, the input parameters comprise temperature, strain rate and physical derivative characteristics, and the output parameters comprise stress; Training a machine learning model by using the training sample set, and optimizing super parameters of the machine learning model by combining a Bayesian optimization algorithm to obtain a prediction model; And acquiring the working condition temperature, the working condition strain and the working condition strain rate of the target metal under the working condition to be detected, inputting the working condition temperature, the working condition strain and the working condition strain rate into the prediction model, obtaining the stress predicted value of the target metal under the working condition to be detected, which is output by the prediction model, and obtaining the stress strain constitutive relation of the target metal under the working condition to be detected.
2. The prediction method of claim 1, wherein obtaining a training sample set comprises: Acquiring real experimental data including temperature, strain rate, stress, and And processing the real experimental data by using a multi-physical coupling constitutive equation to generate the physical derivative characteristics.
3. The prediction method of claim 1 wherein the physically derived features include logarithmic strain rate, normalized reciprocal temperature, and strain hardening rate.
4. The prediction method of claim 2, wherein the equation of the multi-physical coupling constitutive equation is: ; In the formula, The temperature is 25 ℃ reference stress of metal, T is temperature, Q is metal heat activation energy, R is gas constant, H is strain rate-strain coupling coefficient, epsilon is strain; is the strain rate; K is the hardening rate-temperature coupling coefficient; is the strain hardening rate; Is the normalized temperature reciprocal.
5. The prediction method according to claim 4, wherein sigma 0 is obtained by fitting elastic stress under the working conditions of 25 ℃ and 0.001s -1 , Q is obtained by fitting stress-temperature relationship in the range of 800 ℃ to 1200 ℃, and H and K are obtained by fitting experimental data in a plastic stage.
6. The prediction method of claim 1, wherein an optimization formula for optimizing the super-parameters of the machine learning model using a bayesian optimization algorithm is: F obj =0.5×MSE+0.3×MAE+0.2×(1- ); In the formula, ; ; ; Representing the i-th stress value predicted by the predictive model, Representing the experimentally measured ith stress value, Representing the ith stress value generated by the multiple physically coupled constitutive equation, Representing the decision coefficients of the multiple physically coupled constitutive equation.
7. The method of claim 1 or 6, wherein the super parameters include a maximum tree depth of 10-15 layers, a minimum number of samples of leaf nodes of 5-10, and a feature subset dimension of 3-5 dimensions.
8. The prediction method of claim 1, further comprising model verification of the prediction model; Optionally, evaluating the prediction model performance using hierarchical 10-fold cross-validation, calculating a global evaluation index: ; ; Wherein SSE k is the sum of squares of the k-th folding residual errors, SST k is the sum of squares of the k-th folding total errors, and the prediction model meets the following requirements 0.99, MAE global <5 MPa, MSE global <30 MPa 2 .
9. A machine learning based metal stress strain constitutive relation prediction system, comprising: The system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module is used for acquiring a training sample set and working condition temperature, working condition strain and working condition strain rate data of working conditions to be detected; the model construction and optimization module is used for training the machine learning model by utilizing the training sample set and optimizing the super parameters of the machine learning model by combining a Bayesian optimization algorithm to obtain a prediction model; The prediction module is used for inputting the working condition temperature, the working condition strain and the working condition strain rate data of the working condition to be detected into the prediction model to obtain the stress predicted value of the working condition to be detected, which is output by the prediction model, and obtaining the stress-strain constitutive relation under the working condition to be detected.
10. The prediction system of claim 9 further comprising a verification module for model verification of the prediction model.

Description

Metal stress strain constitutive relation prediction method and system based on machine learning Technical Field The invention belongs to the technical field of material mechanical property modeling and machine learning intersection, and particularly relates to a metal stress strain constitutive relation prediction method and system based on machine learning. Background The TA15 titanium alloy is a key material in the aerospace and medical fields by virtue of excellent specific strength, corrosion resistance and biocompatibility. The thermoforming process of the TA15 titanium alloy usually involves high temperature of 800-1200 ℃ and multi-strain rate of 0.001 s -1~1 s-1, and the deformation process comprises three stages of elasticity (predominance of elastic force among atoms), plasticity (dislocation movement and work hardening) and dynamic recrystallization (dislocation density reduction and stress softening), and physical mechanism differences of each stage are obvious, so that the stress-strain constitutive relation presents strong nonlinearity and multi-parameter coupling characteristics. The traditional constitutive model (such as Johnson-Cook and Zerill-Armstrong models) relies on an empirical formula and manual parameter calibration, and has at least the obvious defects that 1) stress softening effect caused by dynamic recrystallization in a high temperature domain (> 800 ℃) is difficult to accurately characterize, errors often exceed 15%, 2) parameter calibration requires a large number of high-temperature and multi-strain rate experiments to be carried out, the period is as long as several months, the cost is high (the cost of a single experiment is over ten thousand yuan), and 3) when the model is extrapolated to extreme working conditions (such as 1200 ℃ and 0.001 s -1), the prediction errors are obviously increased (> 20%) due to lack of physical mechanism support. The existing machine learning modeling method improves the precision through data driving, but has the following defects that 1) the characteristic engineering only depends on original experimental data (strain, strain rate and temperature), and does not incorporate a physical mechanism of material deformation, so that the model has poor physical interpretability and cannot explain the essential reason of stress change at high temperature, and 2) the data enhancement technology mostly adopts unconstrained generation (such as pure random sampling), and the generated sample possibly violates the physical rule (such as pure random sampling)<0 Non-physical softening), instead reducing the model generalization ability, 3) hyper-parametric optimization with fixed objective function (e.g. only minimizing MSE), without considering physical consistency, resulting in model fitting data but departing from the actual deformation law, with insufficient robustness in small sample scenarios. Therefore, a titanium alloy stress-strain prediction method based on a traditional model and an existing machine learning model still needs to be improved. Disclosure of Invention The invention aims to provide a metal stress strain constitutive relation prediction method and system based on machine learning, and the prediction method provided by the invention predicts the precisionAnd the experimental cost is reduced by 80 percent, the method is suitable for the hot forming and mechanical simulation scenes of aerospace structural members, and at least solves the problems of poor physical interpretability and redundancy of machine learning characteristics of the traditional model. The invention provides a metal stress-strain constitutive relation prediction method based on machine learning, which comprises the following steps of obtaining a training sample set, wherein each training sample in the training sample set independently comprises input parameters and output parameter data, the input parameters comprise temperature, strain rate and physical derivative characteristics, the output parameters comprise stress, training a machine learning model by utilizing the training sample set, optimizing super-parameters of the machine learning model by combining a Bayesian optimization algorithm to obtain a prediction model, obtaining working condition temperature, working condition strain and working condition strain rate of target metal under a working condition to be detected, inputting the working condition temperature, working condition strain and working condition strain rate of the target metal under the working condition to be detected into the prediction model, and obtaining a stress predicted value of the target metal under the working condition to be detected, which is output by the prediction model, so as to obtain the stress-strain constitutive relation of the target metal under the working condition to be detected. In some embodiments, obtaining the training sample set includes obtaining real experimental data including temperature, strain rate, stress,