Phase discovery with active learning: Application to structural phase transitions in equiatomic NiTi

Nickel titanium (NiTi) is a protypical shape-memory alloy used in a range ofbiomedical and engineering devices, but direct molecular dynamics simulationsof the martensitic B19' -> B2 phase transition driving its shape-memorybehavior are rare and have relied on classical force fields with limitedaccuracy. Here, we train four machine-learned force fields for equiatomic NiTibased on the LDA, PBE, PBEsol, and SCAN DFT functionals. The models are trainedon the fly during NPT molecular dynamics, with DFT calculations and modelupdates performed automatically whenever the uncertainty of a local energyprediction exceeds a chosen threshold. The models achieve accuracies of 1-2meV/atom during training and are shown to closely track DFT predictions of B2and B19' elastic constants and phonon frequencies. Surprisingly, in large-scalemolecular dynamics simulations, only the SCAN model predicts a reversible B19'-> B2 phase transition, with the LDA, PBE, and PBEsol models predicting areversible transition to a previously uncharacterized low-volume phase, whichwe hypothesize to be a new stable high-pressure phase. We examine the structureof the new phase and estimate its stability on the temperature-pressure phasediagram. This work establishes an automated active learning protocol forstudying displacive transformations, reveals important differences between DFTfunctionals that can only be detected in large-scale simulations, provides anaccurate force field for NiTi, and identifies a new phase.