Machine Learning Enhanced Dynamic Response Modelling of Superelastic Shape Memory Alloy Wires.

Abstract

Superelastic shape memory alloy (SMA) wires exhibit superb hysteretic energy dissipation and deformation capabilities. Therefore, they are increasingly used for the vibration control of civil engineering structures. The efficient design of SMA-based control devices requires accurate material models. However, the thermodynamically coupled SMA behavior is highly sensitive to strain rate. For an accurate modelling of the material behavior, a wide range of parameters needs to be determined by experiments, where the identification of thermodynamic parameters is particularly challenging due to required technical instruments and expert knowledge. For an efficient identification of thermodynamic parameters, this study proposes a machine-learning-based approach, which was specifically designed considering the dynamic SMA behavior. For this purpose, a feedforward artificial neural network (ANN) architecture was developed. For the generation of training data, a macroscopic constitutive SMA model was adapted considering strain rate effects. After training, the ANN can identify the searched model parameters from cyclic tensile stress–strain tests. The proposed approach is applied on superelastic SMA wires and validated by experiments.