在面向飞行器设计的数值模拟中，网格平滑方法是提升前处理流程中网格质量，减少模拟误差的重要手段。传统的优化式平滑方法受限于复杂的迭代求解过程，存在内存开销大、计算效率低等问题。为解决该问题，已有的智能化平滑方法采用神经网络拟合平滑过程，能够实现平滑效率和质量的平衡。然而，已有的方法在应用到三维表面网格时通常采用投影操作或有监督学习来保证网格点的贴体性，引入了额外的计算或数据生成开销。本研究基于无监督学习技术和局部曲面拟合，搭建了面向飞行器表面网格的智能化平滑代理模型GMSNet3D。模型设计了面向表面网格平滑的无监督损失函数，实现了无需高质量监督数据下的智能训练；模型还创新性地引入局部曲面坐标变换来保证平滑后网格点的贴体性。实验结果证明，GMSNet3D模型采用的局部曲面坐标变换方法相比于已有方法中的投影操作实现了13.82倍的加速比；同时，GMSNet3D模型在保证网格平滑质量的同时，与传统的优化式平滑方法相比实现了29.81倍的优化效率提升。

In numerical simulations for aircraft design, mesh smoothing methods are crucial for enhancing mesh quality in the pre-processing stage and reducing simulation errors. Traditional optimization-based smoothing methods are limited by com-plex iterative solving processes, leading to high memory consumption and low computational efficiency. To address these issues, existing intelligent smoothing methods use neural networks to learn the smoothing process, achieving a balance between smoothing efficiency and quality. However, when applied to three-dimensional surface meshes, these methods often rely on projection operations or supervised learning to ensure mesh node conformity, which introduces additional computational or data generation overhead. This study develops an intelligent smoothing surrogate model, GMSNet3D, specifically designed for aircraft surface meshes, based on unsupervised learning techniques and local surface fitting. The model incorporates an unsupervised loss function tailored for surface mesh smoothing, enabling intelligent training without the need for high-quality supervised data. Furthermore, the model innovatively introduces local surface coordinate transformation to ensure the conformity of smoothed mesh nodes. Experimental results demonstrate that the local sur-face coordinate transformation method used in the GMSNet3D model achieves a speedup of 13.82 times compared to projection operations in existing methods. Additionally, while maintaining mesh smoothing quality, GMSNet3D achieves a 29.81-fold improvement in optimization efficiency compared to traditional optimization-based smoothing methods.