Investigating the use of terrain-following coordinates in AI-driven precipitation forecasts
Yingkai Sha †, John S. Schreck †, William Chapman ‡, David John Gagne II † · 2025
Read Investigating the use of terrain-following coordinates in AI-driven precipitation forecasts by Yingkai Sha †, John S. Schreck †, William Chapman ‡, David John Gagne II † (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-00332.txt.
Introduction
Yingkai Sha and co-authors add terrain shaped coordinates to AI weather models in a 2025 study. AI forecasts tend to blur rain, above all missing downpours, which limits their worth for flood watch.
A test model rebuilt on a one degree grid with the new coordinates sharpens rain calls a lot. Extreme events and the spread of rain strength both score better. The coordinates pair well with rules that keep mass and energy honest, cutting a nagging drizzle bias. Case checks show near ground winds drawn more truly, which helps the model learn how rain links to the air.
Catalog record
| CREATOR | Yingkai Sha †, John S. Schreck †, William Chapman ‡, David John Gagne II † |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-00332
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Investigating the use of terrain-following coordinates in AI-driven precipitation forecasts Thanks: Citation : Sha, Y. et al. Investigating the contribution of terrain-following coordinates and conservation schemes in AI-driven precipitation forecasts. Submitted for publication in Geophysical Research Letters Yingkai Sha †, John S. Schreck †, William Chapman ‡, David John Gagne II † Abstract Artificial Intelligence (AI) weather prediction (AIWP) models often produce “blurry” precipitation forecasts. This study presents a novel solution to tackle this problem—integrating terrain-following coordinates into AIWP models. Forecast experiments are conducted to evaluate the effectiveness of terrain-following coordinates using FuXi, an example AIWP model, adapted to 1.0 ∘ grid spacing data. Verification results show a largely improved estimation of extreme events and precipitation intensity spectra. Terrain-following coordinates are also found to collaborate well with global mass and energy conservation constraints, with a clear reduction of drizzle bias. Case studies reveal that terrain-following coordinates can represent near-surface winds better, which helps AIWP models in learning the relationships between precipitation and other prognostic variables. The result of this study suggests that terrain-following coordinates are worth considering for AIWP models in producing more accurate precipitation forecasts. Plain Language Summary Artificial intelligence (AI) weather prediction (AIWP) models can forecast precipitation but often produce “blurry” results—overestimating drizzle while underestimating heavy rainfall. We aim to reduce this blurriness problem by developing an AIWP model using a hybrid sigma-pressure vertical coordinate system that follows the terrain, rather than pressure coordinates that intersect with terrain. We also developed and applied physical constraints under this coordinate system. We compare the results from this new AIWP model with its baseline versions and find that the new version can make more accurate precipitation forecasts with an improved distribution of precipitation values from light to heavy precipitation events. We suggest that it is beneficial for AIWP models to incorporate terrain-following coordinates in producing more accurate precipitation forecasts. 1 Introduction Artificial Intelligence (AI) weather prediction (hereafter AIWP) models have achieved great success, driven by advances in neural network architectures, computational devices, and the availability of high-quality reanalysis datasets, such as the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis version 5 ( Hersbach et al., (2020) , ERA5; ) . State-of-the-art AIWP models can generate forecasts with low computational costs. Their forecast skill scores are competitive with top global Numerical Weather Prediction (NWP) models and can be improved further with additional training data ( Bi et al., (2023) ; Lam et al., (2023) ; Chen et al., (2023) ; Nguyen, Brandstetter et al., (2023) ; Bonev et al., (2023) ; Bodnar et al., (2024) ; Nguyen, Shah et al., (2023) ; Willard et al., (2024) ; Lang et al., (2024) ; Schreck et al., (2024) , e.g.) . However, AIWP models exhibit lower fidelity in forecasting precipitation compared to prognostic, free-atmosphere variables. Recent evaluations on Weatherbench2 reveal that AIWP models tend to produce “blurry” precipitation forecasts. In many cases, they are verified to be worse than the ECMWF Integrated Forecast System (IFS) Rasp et al. ((2024)) . To tackle the problem of AI-driven precipitation forecasts, this study proposes a novel solution—integrating terrain-following coordinates into AIWP models. Terrain-following coordinates have been used in NWP models since Phillips ((1957)) . Among these, hybrid sigma-pressure coordinates– which follow terrain near the surface and transition to constant pressure levels aloft –have seen the widest adoption Simmons & Burridge ((1981)) ; Ritchie ((1991)) ; European Centre for Medium-Range Weather Forecasts ((2016)) . As discussed in Rasp et al. ((2024)) , an important reason for the blurry AI-driven precipitation forecasts is the lack of representation of local-scale features. Here, we hypothesize that hybrid sigma-pressure coordinates have the potential to represent near-surface weather conditions better and help AIWP models understand the relationships between local-scale information and precipitation characteristics when making forecasts. Several studies have identified the effectiveness of using physical constraints in AIWP models ( Wang et al., (2025) ; Watt-Meyer et al., (2023) ; Yuval et al., (2024) ; Sha et al., (2025) , e.g.) . Many of which also benefit precipitation forecasts. Notably, in Sha et al. ((2025)) , the integration of global mass and energy conservation constraints in constant pressure level AIWP models provided a clear reduction of drizzle bias—a common problem of weather models ( Gutowski Jr et al., (2003) , e.g.) . To ensure that the benefits of physical constraints can be preserved under terrain-following coordinates, in this study, we developed a new set of conservation schemes as constraints, similar to that of Sha et al. ((2025)) , but specifically for hybrid sigma-pressure levels to examine its interaction with the terrain-following coordinates. We evaluate the proposed solutions above using FuXi Chen et al. ((2023)) , an AIWP model architecture validated on Weatherbench2 Rasp et al. ((2024)) . In our experiments, the ERA5, re-gridded to 1.0 ∘ grid spacing, is used for training, while the Integrated Multi-satellitE Retrievals for Global Precipitation Measurement ( Huffman et al., (2020) , IMERG; ) serves as the verification target. The technical foundation of this study is based on the Community Research Earth Digital Intelligence Twin ( Schreck et al., (2024) , CREDIT; ) platform hosted at NSF National Center for Atmospheric Research (NCAR). This study addresses three key questions: (1) Can terrain-following coordinates improve AI-driven precipitation forecasts? (2) Can AIWP physical constraints be applied under terrain-following coordinates and provide combined precipitation forecast benefits? (3) Can we provide a physical basis for these improvements? By answering the questions, we aim to bring insights into the limitations of constant pressure level AIWP models in precipitation forecasts and how such limitations can be tackled. In addition, we hope to inspire future innovations to introduce more atmospheric science domain knowledge to the development of AIWP models. 2 Data and methods 2.1 Data Table 1: Input and output variables of the AIWP models. Type Variable Name Units Role Upper air a Zonal Wind \mathrm{m\cdot s^{-1}} Prognostic, Instantaneous Meridional Wind \mathrm{m\cdot s^{-1}} Air Temperature \mathrm{K} Specific Total Water b \mathrm{kg\cdot kg^{-1}} Geopotential c \mathrm{m^{2}\cdot s^{-2}} Single level Mean Sea Level Pressure d \mathrm{Pa} Prognostic, Instantaneous Surface Pressure e \mathrm{Pa} 2-Meter Temperature \mathrm{K} 10-Meter Zonal Wind \mathrm{m\cdot s^{-1}} 10-Meter Meridional Wind \mathrm{m\cdot s^{-1}} Flux form f Total Precipitation \mathrm{m} Diagnostic, Cumulative Evaporation \mathrm{m} Top-of-atmosphere Net Solar Radiation \mathrm{J\cdot m^{-2}} Outgoing Longwave Radiation \mathrm{J\cdot m^{-2}} Surface Net Solar Radiation \mathrm{J\cdot m^{-2}} Surface Net Longwave Radiation \mathrm{J\cdot m^{-2}} Surface Net Sensible Heat Flux \mathrm{J\cdot m^{-2}} Surface Net Latent Heat Flux \mathrm{J\cdot m^{-2}} Top-of-atmosphere Incident Solar Radiation \mathrm{J\cdot m^{-2}} Input-only, Cumulative Others Sea-ice Cover n/a Input-only, Instantaneous Geopotential at the Surface \mathrm{m^{2}\cdot s^{-2}} Input-only, Static Land-sea Mask n/a Input-only, Static Soil Type n/a Input-only, Static a Upper air variables are prepared separately on constant pressure levels and hybrid sigma-pressure levels. b Specific total water is the combination of specific humidity, cloud liquid water content, and rainwater content. c Geopotential is available for constant pressure level models only. d Mean sea level pressure is available for constant pressure level models only. e Surface pressure is available for hybrid sigma-pressure level models only. f Flux form variables are accumulated every 6 hours. Downward flux is positive. In this study, the AIWP models are trained using the ERA5. Two sets of upper-air variables are prepared: one on hybrid sigma-pressure levels