Optimizing feasible domains for parameter identification in multi-scales constitutive model of La-TZM alloy

This paper focuses on La-TZM alloy and introduces a novel method for multi-scale constitutive model parameter identification based on the BP neural network (BPNN) which has not been reported yet, aiming to enhance model prediction accuracy and improve parameter identification efficiency. When the Genetic Algorithm (GA) is used to identify parameters in multi-scale constitutive models with internal variables, common issues such as unstable results and limited prediction accuracy of micro-scale variables arise, and solutions to these problems have not been reported. Initially, traditional GA is used for parameter identification, and a parameter evaluation method based on parameter perturbation experiments is proposed to analyze the sensitivity of parameters to the model prediction results. Subsequently, BPNN is utilized to refine the threshold range of parameter identification. The results indicate that the multi-scale constitutive model of La-TZM alloy includes nine sensitive parameters (ΔG, α_d, ρ, Q_b, β₁, γ₁, n₁,n₂,n₃, and f_g). After optimizing the threshold values using BPNN, the stability of model identification and the prediction accuracy of micro-scale variables were significantly improved. Notably, the calculated results of the internal variable ρ_gd were consistent with microstructural test data, which can represent the evolution of geometrically necessary dislocations density related to La₂O₃ second-phase particle. In summary, this study proposes a new method for identifying multi-scale constitutive model parameters, which can enhance the macro and micro characterization accuracy of material plastic deformation.