This paper proposes an unsupervised mesh movement network with a novel M-Uniform loss, achieving generalization across diverse PDEs and mesh geometries without labeled data.