Hybrid identification method of coupled viscoplastic-damage constitutive parameters based on BP neural network and genetic algorithm

• Published in: Engineering fracture mechanics 2021-11, Vol.257, p.108027, Article 108027
• Main Authors: Yao, Dan, Duan, Yong-chuan, Li, Mu-yu, Guan, Ying-ping
• Format: Article
• Language: English
• Subjects: AA6061 alloy; Aluminum base alloys; Back propagation networks; BP neural network; Calibration; Constitutive model; Constitutive models; Damage; Genetic algorithm; Genetic algorithms; Identification methods; Mathematical models; Metal sheets; Neural networks; Optimization; Parameter identification; Thermal deformation behavior
• ISSN: 0013-7944; 1873-7315
• DOI: 10.1016/j.engfracmech.2021.108027

•A coupled viscoplastic-damage constitutive model of AA6061 including the thermal deformation and damage evolution was established.•The neural network structures for predicting the constitutive parameters of viscoplastic and damage were determined.•A hybrid identification strategy for calibrating the parameters of the constitutive model based on the BP neural network and genetic algorithm was established.•The parameters in the coupled viscoplastic-damage constitutive model of the 6061 and 6016 aluminum alloys were identified using the hybrid identification method. The constitutive model based on the theoretical framework of coupled viscoplastic-damage involves calibration of multiple and high coupling parameters. The inverse calibration by genetic algorithm (GA) with global search ability has some challenges as the dependence on the selection of the initial population, massive computation, and convergence inconsistency. To obtain statistical knowledge from state data to avoid subjective experience, a hybrid identification method based on the BP neural network and GA is proposed. A coupled viscoplastic-damage constitutive model based on the thermal deformation and microstructure evolution is established. The parameters in the model are determined based on the hybrid identification method. Two types of aluminum alloy sheets are selected to test the generalization, and mean square errors (RMSE) are 2.46 and 4.89, respectively. The results indicate that this method has higher accuracy than the inverse calibration method based on single optimization algorithm.