This project demonstrates flow-field reconstruction using diffusion-based methods. From left to right, the visualization shows the input field, the reconstructed output, and the ground-truth flow.
To assess generalization to out-of-distribution data, the model was evaluated on 256×256 tasks using only 1.5625% of the input information.
Shu, D., Li, Z., & Farimani, A. B. A physics-informed diffusion model for high-fidelity flow field reconstruction. arXiv:2211.14680v2, 2023. https://arxiv.org/abs/2211.14680