A new model for generating temporal super-resolution of 4D scientific simulation data

Abstract
This study presents a novel deep learning-based temporal super-resolution (TSR) model for time-varying volumetric data, addressing the challenge in large-scale spatiotemporal simulations: the inability to store complete simulation results due to hardware limitations in I/O speed and storage capacity, which forces researchers to retain only sparse time steps and compromises the fidelity of downstream analysis. The aim is to enhance time-sparsed data by generating refined intermediate time steps, thereby enabling robust scientific analysis compromised by incomplete data storage. The proposed model IVA-TSR employs a generator-discriminator architecture: the generator synthesizes intermediate time steps by integrating multi-scale convolutional layers and self-attention to capture both spatial features (local geometric details and global structural relationships) and complex non-linear temporal dynamics between sparse simulation steps, while the discriminator ensures structural consistency in spatial fidelity between synthesized and ground-truth volumes through adversarial training. The model was applied to various datasets and compared with linear interpolation (LERP), TSR-TVD, and recurrent neural network (RNN) methods. Results show that IVA-TSR achieves higher PSNR, SSIM, and lower LPIPS values, indicating superior performance in generating time-resolved sequences. In conclusion, IVA-TSR provides a robust solution for enhancing the temporal resolution of time-varying data, outperforming existing methods and enabling more effective analysis and visualization.