Extended Data Table 1 Configuration of various flow cases in the synthetic dataset

From: Recurrent graph optimal transport for learning 3D flow motion in particle tracking

  1. We generate a training dataset with more than 16000 samples, each of which contains two particle sets and the corresponding ground-truth flow. The synthetic dataset consists of six kinds of flow scenarios generated by using computational fluid dynamics (CFD), where some of them are extracted from the Johns Hopkins Turbulence Databases. Here, we list the name of flow motion patterns, the range of observation volume, the range of duration time for particle trajectories in the simulation and the range of displacement ratio ρ. By randomly selecting the parameters in the ranges for each data item, we generate a dataset mimicking diverse conditions in PTV experiments. In addition, a normalization of the particle positions in \({{{\mathcal{P}}}}\) and \({{{\mathcal{Q}}}}\) is performed during training. The domains of the particle sets are scaled to the [0, 2π] × [0, 2π] × [0, 2π] right before being fed into the neural network. The flows (labels) are also scaled accordingly.
Back to article page