EI、Scopus 收录
中文核心期刊
周瑞, 熊宇凯, 储节磊, 阚前华, 康国政, 张旭. 基于机器学习和遗传算法的非局部晶体塑性模型参数识别. 力学学报, 2024, 56(3): 1-12. DOI: 10.6052/0459-1879-23-479
引用本文: 周瑞, 熊宇凯, 储节磊, 阚前华, 康国政, 张旭. 基于机器学习和遗传算法的非局部晶体塑性模型参数识别. 力学学报, 2024, 56(3): 1-12. DOI: 10.6052/0459-1879-23-479
Zhou Rui, Xiong Yukai, Chu Jielei, Kan Qianhua, Kang Guozheng, Zhang Xu. Parameter identification of nonlocal crystal plastic model based on machine learning and genetic algorithm. Chinese Journal of Theoretical and Applied Mechanics, 2024, 56(3): 1-12. DOI: 10.6052/0459-1879-23-479
Citation: Zhou Rui, Xiong Yukai, Chu Jielei, Kan Qianhua, Kang Guozheng, Zhang Xu. Parameter identification of nonlocal crystal plastic model based on machine learning and genetic algorithm. Chinese Journal of Theoretical and Applied Mechanics, 2024, 56(3): 1-12. DOI: 10.6052/0459-1879-23-479

基于机器学习和遗传算法的非局部晶体塑性模型参数识别

PARAMETER IDENTIFICATION OF NONLOCAL CRYSTAL PLASTIC MODEL BASED ON MACHINE LEARNING AND GENETIC ALGORITHM

  • 摘要: 非局部晶体塑性模型考虑了由非均匀变形引起的位错在空间上的重排, 使得其本构模型变得复杂, 可调节参数众多, 因此采用常规的“试错法”难以准确确定这些参数. 虽然遗传算法能够稳健地全局优化解决参数确定问题, 但对于非局部晶体塑性模型, 其计算成本相对较高. 为解决这一问题, 提出了一种耦合机器学习模型的遗传算法, 以有效降低计算成本. 针对含有冷却孔的镍基高温合金的拉伸响应问题, 以单拉应力−应变曲线为目标, 基于屈服应力和最终应力建立评价公式, 使得优化结果与实验尽可能接近. 在这一方法中, 机器学习模型能够通过非局部晶体塑性模型的参数来预测相应的应力值, 从而替代了遗传算法中原本需要的有限元计算过程. 为了分析本构模型参数对单拉力学响应的影响, 研究采用SHAP框架, 并通过有限元结果进行验证. 结果表明, 通过该方法可以有效获取非局部晶体塑性模型参数, 使得参数计算得到的应力−应变响应与实验结果吻合较好. 此外, SHAP框架能够提供本构模型参数的重要程度分析, 以及对屈服应力和最终应力的影响.

     

    Abstract: Nonlocal crystal plasticity models account for the spatial rearrangement of dislocations due to non-uniform deformation, resulting in complex constitutive models with many adjustable parameters. Hence, it is challenging to determine these parameters accurately using the conventional trial-and-error method. Although genetic algorithms can solve the parameter identification problem robustly by global optimization, their computational costs are relatively high for nonlocal crystal plasticity models. To address this issue, this paper proposes a genetic algorithm coupled with a machine learning model to effectively reduce the computational cost. Focusing on the tensile response of nickel-based superalloys with cooling holes, the uniaxial tensile stress-strain curve is the objective, and an evaluation formula is established based on the yield stress and the final stress, to make the optimization result as close as possible to the experiment. Through this method, the machine learning model can predict the corresponding stress values based on the parameters of the nonlocal crystal plasticity model, thus replacing the finite element calculation that was originally required in the genetic algorithm. To analyze the influence of the constitutive model parameters on the uniaxial tensile mechanical response, the SHAP framework is adopted and verified by the finite element results. The results show that the nonlocal crystal plasticity model parameters can be determined effectively by this method, and the stress-strain responses calculated by the parameters agree well with the experimental results. Moreover, the SHAP framework can offer insights into the significance of the constitutive model parameters, as well as their influence on the yield stress and the ultimate stress.

     

/

返回文章
返回