Underground infrastructure faces critical maintenance challenges due to inaccessibility, aging deterioration, and high inspection costs. Digital Twin (DT) technology offers a transformative solution by creating virtual replicas that enable predictive maintenance through services such as rare scenario simulation, virtual sensing interpretation, and uncertainty quantification. Ground Penetrating Radar (GPR) serves as a key non-destructive sensing technology bridging physical and digital twins for underground structures. However, GPR simulation using Finite-Difference Time-Domain (FDTD) methods requires substantial computational resources, creating bottlenecks for efficient DT applications requiring hundreds to thousands of forward simulations. This study presents a lightweight dual-encoder U-Net architecture designed as a rapid surrogate for FDTD-based GPR simulations. With only 3 million parameters, a 13-fold reduction compared to original U-Net, the network costs 5.32 milliseconds for each GPR B-scan simulation, versus 127.46 seconds per B-scan with FDTD simulation using the same GPU configuration. The dual-encoder architecture processes complex electromagnetic properties through separate pathways, with cross-attention fusion enabling robust performance across varying complexity levels. Evaluated on two datasets generated via gprMax, an open-source FDTD-based GPR simulator with performance matrix Structural Similarity Index Measure (SSIM), and Mean Squared Error (MSE), the proposed network demonstrates strong performance on single-void tunnel scenarios (SSIM=0.83, MSE=0.0028) and maintains practical utility on multi-layer reinforced tunnel structures with irregular defects (SSIM = 0.72, MSE=0.0030). By accelerating GPR simulation from minutes to milliseconds, the proposed surrogate model enables previously infeasible capabilities including real-time scenario exploration, interactive survey planning, large-scale synthetic dataset generation, and uncertainty quantification.