Data-Driven Exploration of Defects in Amorphous Oxides: Linking Empirical and Machine Learning Force Fields through Al₂O₃

For future integrated circuit technologies, adopting backend-of-line compatible gate dielectrics (e.g., amorphous Al₂O₃ or HfO₂) and non-silicon semiconductors (e.g., Ga₂O₃ or In₂O₃) have to face new challenges to address the defect physics of amorphous oxides such as Al₂O₃, Ga₂O₃, In₂O₃, and HfO₂. This study focuses on amorphous Al₂O₃ based on first-principles methods, empirical force fields, and machine-learning force fields to model realistic defects. Benchmark amorphous Al₂O₃ structures were generated via melting and with careful convergence on quenching rates as slow as 0.1 K/ps, achieving above-ground-state energies of ~0.15 eV/atom and coordination defect distributions distinct from previous results. For machine-learning forcefields (MLFF), we show that MLFF atomic local energies, although not being a physical observables, can serve as effective descriptors for defect formation energies, when such quantities are not easily defined for defects in amorphous oxides. In the case of amorphous Al₂O₃, we show that local energies correctly identify defects such as undercoordinated oxygen atoms, without the need of structural identification protocols.