Systematic study was conducted on the variation regularity of stress-strain curve, feature point stress, dissipated energy and equivalent damping ratio of shape memory alloy (SMA) wires changed with wire diameter, strain amplitude, loading rate and loading cyclic number. By nonlinearly modeling experimental results for SMA using the neural network intelligent algorithm (a neural network algorithm with back-propagation training) and optimizing the initial weight and threshold value of neurons using genetic algorithm, a new BP neural network constitutive model for SMA optimized with genetic algorithm is established. This model successfully overcomes the shortcomings of other mathematical models such as the phenomenological Brinson, by which the various influence factors to mechanical properties in an experiment for SMA are hardly simulated exactly. In fact, the average error between experimental and simulated results is only 1.13% by using this model, much better than conventional BP neural network models. The results show that the BP neural networks constitutive model optimized with genetic algorithm can not only predict accurately the superelastic performance of SMA under cyclic loading, but also avoid the no convergence problem caused by concussion of BP network due to the improper initial weight and threshold value set up. Furthermore, this model would be a better model than others because of fully considering the dynamic influence of loading/unloading rate on SMA experiments.
