Abstract: Early prediction of landslide hazards and analysis of internal deformation mechanisms are critical for disaster warning and engineering mitigation. While grounded in physical mechanisms, traditional finite element simulations are labor-intensive and poorly suited for the dynamic demands of complex slope engineering. Empirical models, though efficient, lack accuracy in capturing internal deformation dynamics. To address this gap, we propose a Deep Hierarchical Point Cloud Variational AutoEncoder (DHPC-VAE) that fuses physical priors from finite element simulations with the nonlinear modeling capabilities of deep learning. By integrating surface and internal monitoring data for physics-guided fine-tuning, our model performs progressive, layer-wise reconstruction of 3D deformation fields from surface to subsurface. Case studies demonstrate that DHPC-VAE accurately replicates internal deformation patterns, achieving a maximum correlation coefficient of 0.94 with monitoring data. Further, incorporating InSAR surface deformation enables internal inference with temporal and spatial correlation coefficients up to 0.81 and 0.87, respectively. This framework provides a fully automated, cost-effective solution for internal deformation inversion, bridging model- and data-driven paradigms with physics-informed learning. It holds strong potential for landslide digital twin modeling, intelligent slope monitoring, and real-time early warning applications.