Parameter identification of elastoplastic model for CuCrZr alloy by the neural network-aided Bayesian inference

The constitutive model serves as the foundation for executing structure analysis to obtain the deformation and stress/strain. In this paper, a neural network-assisted Bayesian parameter identification framework is presented to calibrate parameters of the constitutive model while considering the unavoidable uncertainties. The low-cycle fatigue test of the CuCrZr alloy at 700K is first performed to provide realistic data. The posterior distributions are obtained by applying the transitional Markov Chain Monte Carlo method. To accelerate the identification, the neural network is adopted to directly predict the likelihood function value given material parameters. The effect of prior distributions on the identification parameters is also studied. The characteristic parameters of the normal distribution have almost no effect on the identification results. In the absence of prior information, uniform prior distributions can be used to perform Bayesian identification of material parameters, and satisfactory identification parameters can also be acquired.