Baselines
The table below provides inference performance and baseline results on all LagrangeBench datasets. Runtimes are evaluated Nvidia A6000 48GB GPU.
Note
Result discussion and hyperparams can be found in the full paper “LagrangeBench: A Lagrangian Fluid Mechanics Benchmarking Suite”.
| Model | #Params | Forward [ms] | MSE5 | MSE20 | |
|---|---|---|---|---|---|
| TGV 2D (2.5K) | GNS-5-64 | 161K | 1.4 | 6.4e-7 | 9.6e-6 |
| GNS-10-128 | 1.2M | 5.3 | 3.9e-7 | 6.6e-6 | |
| SEGNN-5-64 | 183K | 9.8 | 3.8e-7 | 6.5e-6 | |
| SEGNN-10-64 | 360K | 20.2 | 2.4e-7 | 4.4e-6 | |
| RPF 2D (3.2K) | GNS-5-64 | 161K | 2.1 | 4.0e-7 | 9.8e-6 |
| GNS-10-128 | 1.2M | 6.7 | 1.1e-7 | 3.3e-6 | |
| SEGNN-5-64 | 183K | 15.1 | 1.3e-7 | 4.0e-6 | |
| SEGNN-10-64 | 360K | 29.7 | 1.3e-7 | 4.0e-6 | |
| EGNN-5-128 | 663K | 60.8 | unstable | unstable | |
| PaiNN-5-128 | 1.0M | 9.1 | 3.0e-6 | 7.2e-5 | |
| LDC 2D (2.7K) | GNS-5-64 | 161K | 1.5 | 2.0e-6 | 1.7e-5 |
| GNS-10-128 | 1.2M | 5.7 | 6.4e-7 | 1.4e-5 | |
| SEGNN-5-64 | 183K | 10.0 | 9.9e-7 | 1.7e-5 | |
| SEGNN-10-64 | 360K | 21.1 | 1.4e-6 | 2.5e-5 | |
| DAM 2D (5.7K) | GNS-5-64 | 161K | 3.8 | 2.1e-6 | 6.3e-5 |
| GNS-10-128 | 1.2M | 11.9 | 1.3e-6 | 3.3e-5 | |
| SEGNN-5-64 | 183K | 28.8 |
2.6e-6 | 1.4e-4 | |
| SEGNN-10-64 | 360K | 59.2 | 1.9e-6 | 1.1e-4 | |
| TGV 3D (8.0K) | GNS-5-64 | 161K | 8.4 | 3.8e-4 | 8.3e-3 |
| GNS-10-128 | 1.2M | 30.5 | 2.1e-4 | 5.8e-3 | |
| SEGNN-5-64 | 183K | 79.4 | 3.1e-4 | 7.7e-3 | |
| SEGNN-10-64 | 360K | 154.3 | 1.7e-4 | 5.2e-3 | |
| RPF 3D (8.0K) | GNS-5-64 | 161K | 8.4 | 1.3e-6 | 5.2e-5 |
| GNS-10-128 | 1.2M | 30.5 | 3.3e-7 | 1.9e-5 | |
| SEGNN-5-64 | 183K | 79.4 | 6.6e-7 | 3.1e-5 | |
| SEGNN-10-64 | 360K | 154.3 | 3.0e-7 | 1.8e-5 | |
| EGNN-5-128 | 663K | 250.7 | unstable | unstable | |
| PaiNN-5-128 | 1.0M | 43.0 | 1.8e-5 | 3.6e-4 | |
| LDC 3D (8.2K) | GNS-5-64 | 161K | 8.6 | 1.7e-6 | 5.7e-5 |
| GNS-10-128 | 1.2M | 32.0 | 7.4e-7 | 4.0e-5 | |
| SEGNN-5-64 | 183K | 81.2 | 1.2e-6 | 4.8e-5 | |
| SEGNN-10-64 | 360K | 161.2 | 9.4e-7 | 4.4e-5 |