Abstract: The Latin hypercube sampling method was used to uniformly sample the Johnson-Cook (J-C) constitutive parameter space for GH4169 alloy from literatures. The stress-strain curve corresponding to the parameter combination was output by finite element simulation. With the simulated stress-strain curve data as input and J-C constitutive parameters as output, the forward neural network (FNN) model, random forest (RF) model and recurrent neural network (RNN) model were used for training, and Bayesian hyperparameter optimization was conducted on the models. The parameter inversion ability of the three models was compared by the determination coefficient R² of the test set. With the stress-strain curve data obtained by testing as input, the corresponding J-C constitutive parameters were output by the trained machine learning model, and then the stress-strain curve was generated by finite element simulation and compared with the test results. The results show that the R² of FNN model, RF model and RNN model on the test set was 0.847, 0.499, 0.741, respectively; FNN model had the best performance, while RF model had the worst performance and was not suitable for the inversion of material constitutive parameters. The compressive stress-strain curves of GH4169 alloy at 850–1 050 ℃ predicted with FNN model inversion J-C constitutive parameters were more consistent with the test curves, and the average relative error was about 11.9% and 2.1% lower than that of the predicted curves obtained with RNN model inversion J-C constitutive parameters and test curve fitting constitutive parameters, respectively, which verified the effectiveness of the machine learning method in the inversion of J-C constitutive parameters of GH4169 alloy.