A Super-Resolution Method for Non-Uniform Mesh Data based on Deep Neural Networks

Mesh data is vital for simulating complex systems in engineering and scientific computation. Higher mesh resolution enhances accuracy and convergence speed but incurs significant computational costs. To address this, we present the Non-Uniform Mesh Super-Resolution Neural Network (NMSRN), a deep learning-based method that reconstructs low-resolution meshes into high-resolution ones, improving data accuracy and optimizing resource usage. NMSRN integrates an autoencoder with an Efficient Sub-Pixel Convolutional Neural Network (ESPCN) for enhanced feature learning and modular training, boosting reusability and adaptability. Comparative experiments on a 2D low-temperature plasma simulation demonstrate NMSRN’s superior accuracy and generalization. Additionally, we investigate fine-tuning strategies for efficiently adapting pre-trained models to new mesh super-resolution tasks.