The deep-learning coarse-grained particle (DCGP) method, a multiscale model combining deep learning and physical principles to incorporate crystalline microstructure, is proposed for metallic materials. In the DCGP method, an atom cluster is equivalent to a particle with additional generalized degrees-of-freedom learned by neural networks. The dynamics governing equations of particles are derived by the energy decomposition and the reduced-order mapping. A point-cloud Transformer autoencoder (PTA) network is developed to project atomic configurations onto low-dimensional internal variables. It adequately utilizes the nonlinear representation capacity of deep learning to enable the characterization of defects within particles. We train distinct neural networks for information transfer and force-field predictions based on the datasets from molecular dynamics (MD), and seamlessly integrate the networks into the particle dynamics framework. Numerical experiments reveal that the DCGP method achieves robust performance across diverse dynamics problems and materials, exhibiting nice consistency with MD reference solutions while delivering a computational speedup of over two orders of magnitude. A micrometer-scale flyer impact simulation is conducted using the DCGP method to demonstrate its ability for analysis at the mesoscopic scale. This study provides an effective approach to enhancing the computational scalability of cross-scale mechanical modelling.