Machine-learned forcefield guided atomistic modelling and two-level systems search at the Al/AlO_x/Al interfaces

Two-level systems (TLS) are a major source of decoherence in Josephson junctions, primarily originating from defects at the Al/AlO_x interface. Atomistic simulations can reveal detailed links between local atomic structure and TLS formation, but accurate Density Functional Theory (DFT) methods are computationally limited in scale.

Here, we employ a combination of our recently developed machine-learned force field and DFT methods to model a series of 1-3 nm thick Al/AlO_x/Al heterostructures of controlled variation in stoichiometry and perform large-scale TLS searches using Activation-Relaxation transition state search methods. Owing to the highly disordered nature of AlO_x films, extensive sampling of diverse interface configurations is required to reliably distill correlations between local chemistry and structural motifs hosting TLS in the oxide. Using this framework, we identify characteristic local motifs correlated with TLS, their spatial extent, and associated dipole moments. Our study aims to correlate different TLS types to structural features arising from different interface growth conditions, offering insights into mitigating decoherence in superconducting qubits.