Generalization of learned and physics based models to different initial conditions. The number next to each video conveys the fraction of solenoidal (as opposed to compressive) component, which ranges from 0 (most compressive) to 1 (most solenoidal). The training data consisted of f=0.66. Simulators shown include the physics-based solver Athena++ at resolutions of 128^3 (ground truth), 64^3, and 32^3 and learned Dilated ResNet model at resolution of 32^3, with and without training noise.

Models generalize reasonably well to more solenoidal but not to more compressive initial conditions, possibly due to faster turbulence decay under compressive conditions.