This study presents a deep learning-based surrogate model utilizing Sparse Convolutional Neural Networks (SCNN) for the rapid and precise prediction of 3D stress fields on parameterized geometries. While conventional dense convolutional architectures are effective for grid-like data, their application to 3D structural analysis is often hindered by extreme memory demands and computational inefficiency, as they perform redundant calculations across the entire spatial domain regardless of occupancy. To overcome these limitations, the proposed framework represents 3D structural geometries as sparse voxel grids and leverages sparse convolution layers to focus computational resources exclusively on occupied voxels. By eliminating unnecessary operations on empty spatial regions, the model achieves high-fidelity field prediction with a significantly reduced memory footprint and superior computational efficiency compared to standard deep learning models. The framework’s performance is validated through numerical examples involving parameterized shell-based structures, such as ship components. The results demonstrate that the SCNN-based model achieves outstanding performance, enabling near-real-time stress evaluation through reduced computational overhead. This approach offers a robust and scalable solution for efficient on-site structural monitoring and high-throughput design optimization processes.