Fig. 1

Pipeline of the proposed model for long-term motion prediction of fluid flows

Table 1

Parameters of U-Net-based autoencoder layers

Fig. 2

Example simulation sequences visualization of our 2D datasets

Table 2

Statistics of simulation datasets used in this paper

Table 3

Statistics of long-term prediction sub-networks

Fig. 3

Time-varying PSNR and cosine similarity of long-term prediction sub-networks

Table 4

Comparison between our long-term velocity prediction sub-network and LSP model on rising smoke datasets

Fig. 4

Time-varying PSNR and cosine similarity of long-term velocity prediction sub-network and LSP model

Fig. 5

Visualization Result comparison between our long-term velocity prediction sub-network and LSP model

Table 5

Statistics of long-term density prediction sub-network

Fig. 6

Long-term density and velocity field prediction of single source smoke datasets (resolution of 64×64)

Fig. 7

Long-term density field prediction of single source smoke with obstacles datasets (resolution of 128×128)

Fig. 8

Long-term density field prediction of rotating cup datasets (resolution of 64×64)

Fig. 9

Long-term density field prediction of rotating and moving cup datasets (resolution of 64×64)

Fig. 10

Long-term density field prediction of single source smoke datasets (resolution of 64×64)

Fig. 11

Long-term density field prediction of 3D single source smoke datasets (resolution of 64×64×64)

Fig. 12

Long-term density field prediction of 3D single source smoke with obstacles datasets (resolution of 64×64×64)

Table 6

Timing of a simulation step computed via long-term velocity prediction sub-network, LSP model, and LSS model averaged over testing data (50 time steps prediction) Our model achieves further speed-ups than the LSP and LSS models (ms)

Table 7

Timing of a simulation step computed via long-term prediction sub-networks averaged over testing data (50 time steps for velocity prediction and 150 time steps for density prediction)