HydroNet FVM-PINN
HydroNet's FVM-PINN neural network, trained per grid with the recipe of its own dam-break example (64×5 network, 5000 Adam epochs).
Research code, run with the settings of its own examples. Ran 2 tests; fewer than 5 with other models, so not ranked on the leaderboard. Score cards from: HydroNet FVM-PINN (0d98ff5).
| Test | Difference, rank | Checks |
|---|---|---|
| Dam break onto dry ground A1F-010 | #20 190 % off of 20 models | • Flood front 713 m ahead of exact; · Water conserved 3.4 % error The trained network spreads 0.3–0.4 m of water over the whole dry channel downstream, so its front is 713 m ahead of the exact one; it gains 3.3 % of the water, and the error grows on finer grids. Its own finite-volume engine, by contrast, never starts the flow. |
| Dam break into shallow water A1F-011 | #20 24 % off of 20 models | • Average depth error 13 cm off; · Water conserved 5.3 % error The trained network smears the bore and the rarefaction: average depth error 12.6 cm, 4.5 % of the water lost, and no improvement on finer grids. Trained with the recipe of HydroNet's own dam-break example, which on its own 20 m, 1 s case gives an RMS error of 6 cm on a 2 m dam (3 %); this case is 2 km over 30 s. |