Learning to Control PDEs with Differentiable Physics
P. Holl, V. Koltun, N. Thuerey, ICLR 2020
Back

Supplemental Material



Below we show selected examples for each experiment in the paper. The directories in the supplemental material contain more examples for each experiment, all of which were randomly selected from the test set.



Figure 5a: Natural fluid flow reconstruction


Each video shows source and target (left), reconstruction (middle), and predictions (right).
Resolution: 128x128, timesteps: 64.



Staggered execution, supervised training



Staggered execution, differentiable physics training



Prediction refinement, differentiable physics training




More prediction refinement examples:






Figure 5b: Shape transitions


Each video shows source and target (left) and reconstruction (right).
Resolution: 128x128, timesteps: 16.



Staggered execution



Prediction refinement


Iterative optimization




More examples: Staggered execution (left), prediction refinement (center) and iterative optimization (right).






Figure 6: Indirect fluid control


The objective is to move as much smoke as possible into the desired bucket at the top.
Resolution: 128x128, timesteps: 16.




More examples





Below we show selected examples for each experiment in the paper. The directories in the supplemental material contain more examples for each experiment, all of which were randomly selected from the test set.