4D water droplet collision datasets

We share the simulated water droplet collision datasets modeled with the Navier-Stokes Cahn-Hilliard equations. This data is used to validate our deep-learning 4D reconstruction algorithm, as reported in "4D-ONIX for reconstructing 3D movies from sparse X-ray projections via deep learning" [1]. The datasets are simulated for two scenarios: reproducible processes and quasi-reproducible processes. We share the training datasets and the ground truth files for both scenarios. The ground truth files are 3D movies of droplet collision simulation. The training datasets are simulated projection pairs for the experiment of X-ray Multi-Projection Imaging (XMPI) [2, 3]. The two projections are 23.8° apart. For the reproducible process, the ground truth file contains a single 3D movie with 75 timestamps. The size of each 3D object is 128×128×128. The training dataset contains 16 XMPI experiments of this dynamical process, measured from 16 random angles. The angles differences among the 16 projection pairs are: 0°,2°,13°,16°,26°,28°,43°,52°,64°,74°,87°,95°,102°,115°,130°,144°. For the quasi-reproducible process, the ground truth file contains 16 simulations of droplet collision with a 10% variance in collision velocities and a 10% variance in droplet size. A single projection pair with random orientation (same as the reproducible process) of the sample was selected from each simulation to form a training dataset. Please refer to our article for the difference between reproducible and quasi-reproducible processes and details about collision simulation and projection generation.References: <br>[1] Zhang, Yuhe, et al. "4D-ONIX: A deep learning approach for reconstructing 3D movies from sparse X-ray projections." <i>arXiv preprint arXiv:2401.09508</i> (2024).<br>[2] Villanueva-Perez, Pablo, et al. "Hard x-ray multi-projection imaging for single-shot approaches." Optica 5.12 (2018): 1521-1524.<br>[3] Villanueva-Perez, Pablo, et al. "Megahertz X-ray Multi-projection imaging." arXiv preprint arXiv:2305.11920 (2023).

我们公开了基于纳维-斯托克斯-钱恩-希利厄德方程（Navier-Stokes Cahn-Hilliard equations）建模的水滴碰撞模拟数据集。本数据集用于验证我们的深度学习4D重建算法，相关研究发表于论文《4D-ONIX：基于深度学习的稀疏X射线投影三维动态序列重建》[1]。
本数据集包含两类模拟场景：可复现过程与准可复现过程。我们公开了两类场景下的训练数据集与真值文件（ground truth files）。
真值文件为水滴碰撞模拟的三维动态序列（3D movies）。训练数据集为X射线多投影成像（X-ray Multi-Projection Imaging, XMPI）实验的模拟投影对[2, 3]。两个投影之间的夹角为23.8°。
针对可复现过程，其真值文件包含一段含75个时间戳的三维动态序列，单帧三维物体的尺寸为128×128×128。该场景下的训练数据集包含16组该动态过程的XMPI实验数据，采集自16个随机角度。16组投影对的角度差分别为：0°、2°、13°、16°、26°、28°、43°、52°、64°、74°、87°、95°、102°、115°、130°、144°。
针对准可复现过程，其真值文件包含16组水滴碰撞模拟结果，碰撞速度与水滴尺寸均带有10%的方差。从每组模拟结果中选取一组随机取向（与可复现过程的取向规则一致）的投影对，构成该场景的训练数据集。
关于可复现过程与准可复现过程的差异，以及碰撞模拟与投影生成的详细细节，请参阅我们的相关论文。
参考文献：
[1] Zhang, Yuhe, et al. "4D-ONIX: A deep learning approach for reconstructing 3D movies from sparse X-ray projections". arXiv预印本arXiv:2401.09508, 2024.
[2] Villanueva-Perez, Pablo, et al. "Hard x-ray multi-projection imaging for single-shot approaches". Optica 5.12, 2018: 1521-1524.
[3] Villanueva-Perez, Pablo, et al. "Megahertz X-ray Multi-projection imaging". arXiv预印本arXiv:2305.11920, 2023.