Developing data reduction tools for large-scale particle datasets is essential to address the DOE Office of Science's network, storage, and computing needs in the exascale era. Unlike generic black-box compression methods, our proposed approach emphasizes preserving the intrinsic structures of particle data, thereby retaining critical kinetic physics with minimal information loss. We consider both static and dynamic particle datasets. The first thrust introduces a novel particle-based distribution representation that naturally bridges conventional particle representations and their moments, accompanied by a unique low-pass filter specifically designed for moment data. The second thrust addresses dynamic datasets through nonlinear, structure-preserving reduction techniques applicable to both dissipative and non-dissipative particle dynamics. Additionally, we will advance asymptotic-preserving model discovery by embedding multiscale structures explicitly into machine learning architectures, focusing especially on multiscale systems and closures for scale bridging. Our methodology will be validated using simulation and experimental datasets from DOE applications. Scalable deployment and open-source release of the resulting software packages are also critical objectives of this project. The outcomes of this project will broadly impact DOE mission areas, including fusion energy, accelerator physics, and nuclear physics.