Optimization of Johnson-Cook constitutive parameters of titanium alloy based on a machine learning-accelerated genetic algorithm

The Johnson-Cook (JC) constitutive model is particularly important for describing the mechanical behavior of titanium alloys under large plastic deformation and high strain rates. To address the challenges of high experimental cost and low efficiency in determining the constitutive parameters of titanium alloys, this study proposes an adaptive iterative method that integrates machine learning (ML), genetic algorithm (GA) and finite element simulation (FEM). The method aims to efficiently and accurately determine the JC constitutive parameters of titanium alloys with specific target properties. The Ti-3Al-8V-6Cr-4Mo-4Zr (TB9) titanium alloy is selected as the research object, with eight key constitutive parameters (A, B, n, C, D₁, D₂, D₃, D₄) as optimization features, and five representative mechanical properties (tensile strength (σ_b), elongation (δ), dynamic compressive strength (σ_c), critical fracture strain (ε) and impact energy (A_KU2)) are set as optimization targets, to validate the effectiveness and efficiency of the algorithm. The algorithm consists of these steps: (1) A multilayer perceptron (MLP) model is trained to construct a nonlinear mapping relationship model (MLP_i) between constitutive parameters and corresponding mechanical properties; (2) During GA optimization, the MLP_i model predicts the mechanical properties for each individual in the population. The parameter set that best matches the target properties is efficiently identified through iterative reproduction and evolution; (3) The optimized parameters are input into the FEM model to simulate mechanical properties, which are then fed back to update the MLP_i model to MLP_{i+1} in a data-driven approach. The results show that the model achieves effective generalization within 10 iterations, with high predictive accuracy on the test set (R²_Test-avg = 0.9632). The difference in optimal fitness between the ML prediction properties and the FEM simulation properties corresponding to the constitutive parameters optimized by GA is significantly reduced (Fitness_{_}ML = 0.023, Fitness_{_}FEM = 0.015), and the relative error between FEM-simulated and experimentally measured properties for the final predicted TB9 constitutive parameters is controlled within 10 %, which verifies the accuracy and reliability of the method. This method exhibits broad applicability and can be extended to the efficient inverse determination and optimization of constitutive parameters for any titanium alloys with a priori target mechanical properties.