Skip to content

PDEBench: 2D Navier–Stokes forecasting

Goal

Compare general structured-grid PDE solvers under long autoregressive rollout.

FNO/F-FNO, PIBERT, UPT, DPOT, Poseidon, PDE-Refiner, BCAT, and a U-Net baseline.

Workflow

navier datasets info pdebench
navier recommend --problem navier_stokes --task surrogate --dimension 2 --mesh structured --temporal autoregressive --top-k 10

Required metrics

Relative L2, RMSE, spectral error, stable horizon, kinetic-energy spectrum, divergence error, wall-clock, and peak memory.

Leakage control

Hold out complete PDE parameter regimes and initial-condition families. Do not tune on the final OOD split.

References

Takamoto et al., PDEBench: An Extensive Benchmark for Scientific Machine Learning, NeurIPS Datasets and Benchmarks, 2022.