<p>Mesh smoothing plays an important role in numerical computations, as the quality of a mesh directly affects the accuracy and efficiency of numerical simulations. Researchers often employ various mesh smoothing techniques to improve mesh quality. Methods based on objective function optimization can effectively enhance mesh quality but require iterative solutions to optimization problems, along with carefully designed objective functions and parameters for specific problems, resulting in high computational costs and significant manual intervention. Heuristic methods, such as the Laplacian method, compute new node coordinates explicitly, offering higher efficiency but limited effectiveness in improving mesh quality. In this paper, we propose a novel mesh smoothing method based on machine learning. We apply Convolutional Neural Networks (CNNs) from deep learning to mesh smoothing, leveraging features such as local receptive field characteristics, max pooling, and online data augmentation to design a direct mapping between low-quality and high-quality mesh node coordinates. The trained neural network model can effectively improve the quality of both uniform and non-uniform triangular and quadrilateral meshes on planar and surface domains. Numerical experiments demonstrate that this method achieves superior single-step prediction results compared to traditional Laplacian methods and optimization-based approaches. The resulting mesh quality is comparable to other methods. Additionally, our approach performs consistently well across meshes with varying degrees of perturbation.</p>