RealPDEBench: cylinder sim-to-real forecasting

Goal

Measure whether simulation pretraining improves forecasting on paired experimental cylinder-wake data.

Data

navier datasets download realpdebench --local-dir ./data/realpdebench --pattern "cylinder/**"

Training paradigms

1. numerical only;
2. real only;
3. numerical pretraining followed by real fine-tuning.

Recommended models

PIBERT, FourierFlow, FNO, Transolver, UPT, DPOT, DeepONet, and DMD.

Metrics

RMSE, MAE, relative L2, R², frequency-domain RMSE, kinetic-energy error, update ratio, and OOD parameter performance.

References

Hu et al., RealPDEBench, ICLR 2026; Chakraborty, Pan & Chen, PIBERT, 2026.