Flow Map Compression and Reconstruction via Multi-Resolution Hash Fusion Neural Network

With the increasing demand for scientific computing and visualization, efficient storage and reconstruction of flow field data have become critical challenges. Conventional flow field compression methods often struggle to balance high compression ratios with accurate data recovery and preservation of flow details. In this work, we propose a neural flow map compression approach based on the fusion of multi-resolution hash encoding and coarseresolution dense grid features. Our method leverages neural networks to efficiently represent the velocity vector field, combining the strengths of multi-resolution hash encoding for local detail and dense grid features for global structure. Furthermore, to address the information loss caused by training with fixed integration steps in previous works, we introduce a novel joint training strategy using flow map samples with multiple integration durations, enhancing the model’s ability to capture dynamic variations in the flow field. Experimental results demonstrate that our approach achieves a compression ratio of 94 times, while maintaining high fidelity in streamline generation and flow field detail reconstruction, significantly outperforming existing methods.