Neural network-aided constitutive modeling of cyclic softening in 2.25Cr–1Mo steel across temperatures and strain amplitudes

Abstract

2.25Cr–1Mo steel is widely used in high-temperature components of nuclear and fossil power plants. Accurate modelling of its cyclic behavior over a wide temperature range is essential for structure integrity assessment. In this study, a Chaboche-type constitutive model is employed to describe the cyclic response of 2.25Cr–1Mo steel under various strain amplitudes and temperatures. The isotropic hardening parameter $Q$, defined as the difference between the initial and the stabilized peak stresses, plays a key role in characterizing cyclic softening. To capture the coupled dependence of $Q$ on strain amplitude and temperature, a physics-constrained neural network was developed. The approach incorporates experimental scatter into the calibration process. The predicted parameters are expressed as logarithmic functions of temperature, enabling smooth interpolation and direct implementation in finite element simulations. The proposed model reproduces the experimental cyclic softening behavior with good accuracy. This framework provides a practical and reliable tool for fatigue and inelastic analysis of high-temperature structural components.