Investigating the contribution of terrain-following coordinates and conservation schemes in AI-driven precipitation forecasts

Yingkai Sha, John S. Schreck, William Chapman, David John Gagne II

NSF National Center for Atmospheric Research, Boulder, Colorado, USA

Abstract

Artificial Intelligence (AI) weather prediction (AIWP) models often produce "blurry" precipitation forecasts that overestimate drizzles and underestimate extremes. This study provides a novel solution to tackle this problem---integrating terrain-following coordinates with global mass and energy conservation schemes into AIWP models. Forecast experiments are conducted to evaluate the effectiveness of this solution using FuXi, an example AIWP model, adapted to 1.0-degree grid spacing data. Verification results show large performance gains. The conservation schemes are found to reduce drizzle bias, whereas using terrain-following coordinates improves the estimation of extreme events and precipitation intensity spectra. Furthermore, a case study revealed that terrain-following coordinates capture near-surface winds better over mountains, offering AIWP models more accurate information on understanding the dynamics of precipitation processes. The proposed solution of this study can benefit a wide range of AIWP models and bring insights into how atmospheric domain knowledge can support the development of AIWP models.

Resources

• NSF NCAR Research Data Archive, ERA5 Reanalysis (0.25 Degree Latitude-Longitude Grid)

  • Note: this repository visits RDA internally, e.g., /glade/campaign/collections/rda/data/d633000/e5.oper.an.pl/197901/

• Google Research, Analysis-Ready, Cloud Optimized (ARCO) ERA5 [link]

• Goddard Earth Science Data and Information Science Center (GES-DISC), NASA, Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG) Final Precipitation L3 daily product (GPM_3IMERGDF) version 7.0 [link]

• IFS-HRES and ERA5 climatology from the Weatherbench2 project [link]

• NSF NCAR MILES Community Research Earth Digital Intelligence Twin (CREDIT) [link]

• The implementation of conservation schemes in CREDIT [link]

Navigation

• Derivations of conservation schemes: Pytorch integration
• Results: TS and SEEPS, quantile-based verification, case study

Model weights

Model name	Weights	Notes
FuXi-sigma-base	link	The 1.0 degree FuXi baseline without conservation schemes
FuXi-sigma-physics	link	The 1.0 degree FuXi run with conservation schemes

Citation

Sha, Y., J. Schreck, W. Chapman, D. J. Gagne II, 2025: Investigating the contribution of terrain-following coordinates and conservation schemes in AI-driven precipitation forecasts. In review: Geophysical Research Letters. Pre-print: https://arxiv.org/abs/2503.00332

@article{sha2025investigating,
  title={Investigating the contribution of terrain-following coordinates and conservation schemes in AI-driven precipitation forecasts},
  author={Sha, Yingkai and Schreck, John S and Chapman, William and Gagne II, David John},
  journal={arXiv preprint arXiv:2503.00332},
  year={2025}
}