Direct prediction of quantum accurate forces for multicomponent systems

Phase-change materials (PCM) are routinely exploited in optical data storage, flexible devices and neuromorphic computing due to the extreme electro-optical contrast between crystalline and amorphous states. However, high-quality force-field models for molecular-dynamics (MD) simulations of PCM are not available, while quantum molecular dynamics simulation is limited by the small size of system. We have developed neural-network (NN) force fields for GeTe and Ge₂Sb₂Te₅, which are trained using atomic forces computed by density functional theory. Radial and angular feature vectors are designed and trained, which feature permutational and translational invariance and rotational covariance of forces. The accuracy of the NN force fields is validated by performing MD simulation involving up to 100,000 atoms and computing multiple structural and dynamical properties.