PDEBench Scientific Machine Learning BenchmarkNew

What it provides

PDEBench bundles long-time-series and steady-state datasets for a broad set of PDEs - advection, Burgers, reaction-diffusion, 2D incompressible NS, shallow water, Darcy flow, damped wave and more - plus baseline models (FNO, U-Net, MLP) and evaluation scripts.

Why it matters

Papers proposing neural operators often evaluate on private splits. PDEBench gives you: (a) standardized train/test splits, (b) fixed metrics, (c) baseline numbers to beat - essential for honest model selection.

Suggested path

1. Download one small dataset (Burgers or reaction-diffusion) via the provided scripts
2. Train the U-Net baseline; reproduce the published error curve
3. Swap in an FNO; observe resolution generalization behavior
4. Read the analysis sections on rollout stability over long horizons

Practical notes

• Data volumes range from MBs to GBs; start with the small set
• Metrics include normalized MSE and relative error - keep them consistent when comparing against your own runs