Abstract: To address geometric instability, cross-scale distortion, and high computational cost in isosurface extrac-tion from large-scale multi-resolution particle simulation data, an adaptive isosurface extraction method with a GPU-parallel framework is proposed. A volume-preserving constraint is introduced into anisotropic kernels to ensure consistent volumetric contributions of particles at different scales, improving geometric stability in multi-resolution scenarios. A normal-consistency-driven adaptive octree is employed for hier-archical sampling, refining complex regions while reducing redundant computation in flat areas. Du-al-grid-based isosurface generation is used to avoid topological cracks caused by multi-resolution spatial partitioning. The entire pipeline, including neighbor search, spatial subdivision, scalar field sampling, and isosurface extraction, is implemented on the GPU. Compared with a narrow-band method, GPU memory consumption is reduced to 7%–10% and mesh storage by 25%–57% on single-resolution datasets, while improved geometric consistency and reconstruction stability are achieved on multi-resolution datasets.