Self-regenerative materials are keys to the development of stable catalysts used under high temperature condition, e.g., three-way catalyst converters in automobiles. Among others, metal nanoparticles supported on perovskite oxides are promising ones. However, little is known about their atomistic details, which are crucial for understanding and development of thermally stable catalysts. Herein, we present a machine-learning-enhanced density functional theory study of Pd x O y nanoparticles supported on a Sr 3 Ti 2 O 7 (001) surface and demonstrate that supported oxidized Pd particles fulfill the conditions for the self-regenerative catalysts. Under the oxidative condition, the solid-solution reaction of Pd with the support is found to be preferable but is limited to the vicinity of the surface. Furthermore, formation of PdO-like clusters enhances their binding strength to the support, inhibiting the agglomeration (sintering) of the Pd x O y nanoparticles. We present detailed thermodynamic and electronic structure analyses and clarify the roles of the oxide support, cluster size of the oxidized metal nanoparticle, and the metal−support interaction. This work may provide a guideline for the rational design of the thermally stable catalyst against sintering