High accuracy neural network interatomic potential for NiTi shape memory alloy

Nickel-titanium (NiTi) shape memory alloys (SMA) are widely used, however simulating the martensitic transformation of NiTi from first principles remains challenging. In this work, we developed a neural network interatomic potential (NNIP) for near-equiatomic Ni-Ti system through active-learning based acquisitions of density functional theory (DFT) training data, which achieves state-of-the-art accuracy. Phonon dispersion and potential-of-mean-force calculations of the temperature-dependent free energy have been carried out. This NNIP predicts temperature-induced, stress-induced, and defect-induced martensitic transformations from atomic simulations, in significant agreement with experiments. The NNIP can directly simulate the superelasticity of NiTi nanowires, providing a tool to guide their design.