Large-scale first-principle simulations of amorphous indium oxide
Amorphous indium oxide (a-In$_2$O$_3$) is a high-electron-mobility semiconductor of central importance in thin-film transistors and a promising photoanode for solar-driven water oxidation. Despite sustained experimental and computational investigations, the structural motifs underlying its unusual transport properties and the existence of O-O peroxide-like bonds within its network have remained unresolved. Here we develop a MACE-based machine-learned interatomic potential trained on first-principles molecular dynamics trajectories and use it to generate and analyze amorphous structures containing up to 5120 atoms, two orders of magnitude larger than those adopted in typical ab initio studies. We find X-ray structure factors in excellent quantitative agreement with experiment and we confirm that In$_2$O$_3$ is a poor glass former, with the likely presence of quasi-crystalline regions in amorphous samples. Our large-scale structural analysis reveals extended chains of edge-sharing InO$_k$ polyhedra providing a concrete structural basis for the high electron mobility of a-In$_2$O$_3$. Our results strongly support the formation of O-O peroxide-like bonds in the amorphous network, with a mean length of 1.5 \AA{}. We show that these bonds introduce localized in-gap states near the conduction band minimum, acting as a source of intrinsic n-type self-doping and enhancing sub-gap optical absorption. These effects are detectable via a distinct Raman feature near 850 cm$^{-1}$ that is absent in the IR spectrum. Overall, our results establish a comprehensive structure-property picture of a-In$_2$O$_3$, provide directly testable experimental predictions, and suggest that controlled amorphization is a viable strategy for improving the photoelectrochemical activity of a-In$_2$O$_3$.