Exploring the energy landscape of aluminas through machine learning interatomic potential

Aluminum oxide (alumina, Al$_2$O$_3$) exists in various structures and has broad industrial applications. While the crystal structure of $\alpha$-Al$_2$O$_3$ is well-established, those of transitional aluminas remain highly debated. In this study, we propose a universal machine learning interatomic potential (MLIP) for aluminas, trained using the neuroevolution potential (NEP) approach. The dataset is constructed through iterative training and farthest point sampling, ensuring the generation of the most representative configurations for an exhaustive sampling of the potential energy surface. The accuracy and generality of the potential are validated through simulations under a wide range of conditions, including high temperatures and pressures. A phase diagram is presented that includes both transitional aluminas and $\alpha$-Al$_2$O$_3$ based on the NEP. We also successfully extrapolate the phase diagram of aluminas under extreme conditions ([0, 4000] K and [0, 200] GPa ranges of temperature and pressure, respectively), while maintaining high accuracy in describing their properties under more moderate conditions. Furthermore, combined with our developed structure search workflow, the NEP provides an evaluation of existing $\gamma$-Al$_2$O$_3$ structure models. The NEP developed in this work enables highly accurate dynamic simulations of various aluminas on larger scales and longer timescales, while also offering new insights into the study of transitional aluminas structures.