We use a recently-developed machine-learned Moment Tensor Potential (MTP) trained on data generated with the density functional theory (DFT) and tailored to amorphous silicon coupled with the Activation-Relaxation Technique nouveau (ARTn) to identify and classify two-level systems (TLS). The samples generated using MTP recover experimental results and provide average structural and dissipative properties similar to those obtained with a modified Stillinger-Weber potential, including radial distribution function, defect concentration and internal friction. Atomistic details, however, are significantly different, including the density and type of TLS. In particular, we find that while the density of TLS involving a bond-hopping mechanism is similar for the two potentials, more complex TLSs, such as those involving a Wooten-Winer-Weaire bond exchange, are about twice as common. Analysis also shows that TLSs, for MTP-based models, are mostly isolated and oscillate independently from each other.
@article{arxiv.2511.19746,
title = {An activation-relaxation technique study of two-level system impact on internal dissipation using DFT-based moment tensor potential},
author = {Renaude Girard and Carl Lévesque and Normand Mousseau and François Schiettekatte},
journal= {arXiv preprint arXiv:2511.19746},
year = {2026}
}