<p>The heterogeneity induced by rock mass structural planes significantly affects the accuracy of rock engineering simulations. This research proposes a data-driven enhanced field-scale 3D finite-discrete element method (DF-3DFDEM) that effectively reproduces rock heterogeneity and enables efficient simulation of rock engineering problems. To explicitly represent rock heterogeneity, the upscale element assembly (UEA) embedded with structural planes is generated, where fracture geometry follows probabilistic distributions. A homogenization method is proposed to separately homogenize the continuous and discontinuous deformation decided by the rock matrix and structural planes, establishing the correlation between multi-scale mechanical properties. Multi-scale UEAs are then subjected to various loading paths, producing a mechanical behavior data set considering rock heterogeneity and historical damage through 3D FDEM simulations. To overcome the computational complexity of processing homogenized time series data, a convolutional neural network-based mixture density network (CNN-MDN) is developed. This hybrid deep learning framework integrates dilated convolutional layers to expand receptive fields for historical damage encoding and employs probabilistic output layers to resolve the one-to-many mapping challenges inherent in heterogeneous material responses. Implementation of general-purpose graphics processing unit (GPGPU) achieves computational acceleration through parallel computing. The DF-3DFDEM is validated through the simulation of tunnel excavation at Jinping II Hydropower Station, with results aligning well with field monitoring data. Compared to conventional 3D FDEM with fine elements, it achieves a computational speedup of approximately two orders of magnitude, indicating the computational accuracy and efficiency of the proposed method in rock engineering simulation.</p>