Machine-Learning-Accelerated Particle-Swarm Optimization for New Ground-State Reconstructions on TiO2 (110)

Determining the atomic structure of reduced rutile TiO 2 (110) remains a long-standing challenge due to the complex rearrangements of Ti and O atoms and the vast configurational space accessible under reducing conditions. Here, we employ an advanced global structuresearch framework that integrates deep-potential energy surfaces with particle-swarm optimization, enabling efficient and unbiased exploration across multiple reduction states. Our search identifies a previously unreported Ti 3 O 2 -(2×1) reconstruction that is thermodynamically more stable than the earlier Ti 3 O 2 -(1×2) model. Simulated STM images of this newly identified structure reproduce the experimentally observed rosette motifs with a high degree of similarity. Electronic-structure analysis shows that its enhanced stability originates from strengthened orbital hybridization between the occupied Ti 3+ 3 d states and the O 2 p states. This work refines the atomic-level understanding of a widely studied TiO 2 (110) reconstruction and highlights the capability of machine-learning-accelerated global search to resolve complex oxide surface structures with direct experimental relevance.