Fig. 2.

Learned interpolation (LI) achieves accuracy of direct simulation at $\sim 10\times$ higher resolution. (A) Evolution of predicted vorticity fields for reference (DS 2,048 × 2,048), learned (LI $64\times 64$), and baseline (DS $64\times 64$) solvers, starting from the same initial velocities. The yellow box traces the evolution of a single vortex. (B) Comparison of the vorticity correlation between predicted flows and the reference solution for our model and DNS solvers. (C) Energy spectrum scaled by $k^5$ averaged between time steps 10,000 and 20,000, when all solutions have decorrelated with the reference solution.