Hybrid weather prediction using spectral nudging toward machine-learning forecasts
I. Polichtchouk 1 , M. C. A. Clare 2 , M. Chantry 1 , E. Gascón 2 , M. Maier-Gerber 2 , B. Vanniere 2 , S. Lang 1 · 2026
Read Hybrid weather prediction using spectral nudging toward machine-learning forecasts by I. Polichtchouk 1 , M. C. A. Clare 2 , M. Chantry 1 , E. Gascón 2 , M. Maier-Gerber 2 , B. Vanniere 2 , S. Lang 1 (2026) 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-2604-22522.txt.
Introduction
Polichtchouk and six colleagues test a hybrid forecast that nudges a physics-based weather model toward machine-learned forecasts. Only the large scales of virtual temperature and spin are nudged, aiming to lift large-scale skill while leaving small-scale physics intact. The nudge improves large-scale skill, with gains of up to 1.5 days in the tropics and 12 to 18 hours outside them, and fewer forecast busts. Near-surface extremes hold up, and cyclone tracks gain from better steering flow while storm strength stays close to the physics model.
Catalog record
| CREATOR | I. Polichtchouk 1 , M. C. A. Clare 2 , M. Chantry 1 , E. Gascón 2 , M. Maier-Gerber 2 , B. Vanniere 2 , S. Lang 1 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48992 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.22522
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Hybrid weather prediction using spectral nudging toward machine-learning forecasts I. Polichtchouk 1 , M. C. A. Clare 2 , M. Chantry 1 , E. Gascón 2 , M. Maier-Gerber 2 , B. Vanniere 2 , S. Lang 1 Abstract A hybrid approach to numerical weather prediction is investigated, in which the unperturbed physics-based ECMWF Integrated Forecasting System (IFS) is spectrally nudged toward forecasts from a machine-learned weather forecast model, trained to forecast on model levels. Nudging is applied only to the large scales of virtual temperature and vorticity, with the objective of improving large-scale forecast skill while preserving the dynamical and physical behaviour of the underlying physics-based model at smaller scales. Consistent with previous studies, spectral nudging substantially improves large-scale forecast skill relative to the free-running IFS, with gains of up to 1.5 days in the tropics and 12–18 hours in the extra-tropics, and a reduced frequency of forecast busts. These improvements are achieved while preserving forecast variability. The representation of extreme near-surface weather is maintained or improved. Tropical cyclone track forecasts benefit from improved large-scale steering flow, while storm intensity remains comparable to that of the physics-based model and more physically consistent than in pure machine-learned weather forecast models. These results confirm that scale-selective spectral nudging provides a practical pathway for combining machine-learning and physics-based forecasting systems. 1 Introduction Recent machine-learning weather prediction systems have demonstrated substantial improvements in large-scale forecast skill relative to traditional physics-based numerical weather prediction (NWP) models, particularly when evaluated using metrics such as root-mean-square error and anomaly correlation ( 16 ; 29 ; 4 ; 6 ; 19 ; 20 , e.g.,) . However, deterministic (i.e., unperturbed) machine-learning forecasts are often characterised by smoother spatial structure and reduced small-scale variability. This behaviour arises in part from the use of regression-based loss functions such as mean-squared error, which favour solutions that minimise average error and therefore reduce sensitivity to small-scale displacement errors associated with the so-called “double penalty” effect ( 10 ; 5 , e.g.,) . The autoregressive rollout used during training to learn multi-step forecasts can further contribute to a progressive loss of small-scale variance ( 30 ) . While this smoothing can improve large-scale verification scores, it can also limit the ability of unperturbed machine-learning models to represent phenomena that depend on sharp gradients or localised diabatic processes, including tropical cyclone intensity and orographic precipitation extremes ( 23 ; 8 ; 28 , e.g.,) . One promising strategy for combining the strengths of machine-learning and physics-based forecasting systems is hybrid prediction, in which machine-learning forecasts provide large-scale guidance while the physics-based model retains responsibility for smaller-scale dynamics and physical processes, and for generating the full set of operational forecast products. 15 demonstrated such an approach using spectral nudging in the Environment and Climate Change Canada Global Environmental Multiscale (GEM) model, where large-scale fields were relaxed toward forecasts from the machine-learning model GraphCast. In that configuration, GEM’s virtual temperature and horizontal winds in the 850–250 hPa layer were nudged toward GraphCast forecasts for horizontal wavelengths larger than approximately 2250 km. This hybrid configuration substantially improved large-scale forecast skill relative to the free-running GEM model, yielding gains of up to 34 hours in predictive skill while maintaining realistic small-scale structure and improving tropical cyclone track and precipitation forecasts. A similar approach has also been explored using spectral nudging to couple the CMA-GFS model with forecasts from the FuXi machine-learning system ( 36 ) . Using the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS) as the physics-based model, and a custom-built version of ECMWF’s Artificial Intelligence Forecasting System (AIFS) ( 20 ) as the machine-learning counterpart, we adopt a hybrid approach similar to that of 15 . However, the present implementation differs in several respects. The hybridisation is carried out within the native spectral formulation of the IFS, allowing nudging to be applied directly to prognostic spectral coefficients. In addition, we employ a bespoke AIFS-Single model trained on the full set of 137 IFS model levels. This avoids repeated vertical interpolation between pressure levels and model levels and enables nudging to be applied throughout the troposphere, including near the surface. Finally, the vertical extent of nudging is limited dynamically by the diagnosed lapse-rate tropopause rather than by a fixed pressure layer. In this study we focus on the deterministic (i.e., unperturbed) forecasting framework and examine how spectral nudging toward machine-learning AIFS-Single forecast modifies the behaviour of the IFS. In particular, we assess its impact on large-scale forecast skill, small-scale structure, forecast busts, and the representation of high-impact weather, including tropical cyclones and surface extremes. We also investigate how spectral nudging interacts with ongoing model development. Although the present study focuses on unperturbed forecasting, the hybrid framework also has important implications for ensemble prediction. In principle, nudging could be applied to a member of the probabilistically trained AIFS-ENS system ( 21 ) ; however, the deterministic forecast skill of a single AIFS-ENS member is typically reduced compared to that of AIFS-Single or the unperturbed IFS, and such an approach would therefore not be expected to improve deterministic forecast performance. Extension of the framework to probabilistic forecasting, in which perturbed members are nudged toward AIFS-ENS forecasts, is shown in 31 . This study is structured as follows. Section 2 describes the IFS and AIFS-Single models, along with the spectral nudging methodology. Section 3 presents a brief case study of the extra-tropical wind storm Amy to illustrate the scale-dependent behaviour of the hybrid system. In Section 4 , we evaluate the forecast skill of the hybrid IFS relative to the operational IFS and AIFS-Single using almost two years of 10-day forecasts. Section 5 presents the impact of nudging on high-impact weather, including tropical cyclone track and intensity, as well as extremes of precipitation, 10-m wind speed, and 2-m temperature. Section 6 examines the interaction between spectral nudging and physics-based model development. Finally, Section 7 provides summary, discussion and conclusions. 2 Models and Methods 2.1 ECMWF IFS We use the full-complexity global ECMWF Integrated Forecasting System (IFS), based on cycle 49r1 ( 32 ) , coupled to the 1/4 \degree NEMO ocean model. The IFS employs a semi-implicit, semi-Lagrangian formulation ( 37 ; 14 ; 9 ) and is horizontally discretised using a spherical harmonic spectral expansion, combined with a cubic-octahedral reduced Gaussian grid ( 26 ) . All forecasts in this study are run at TCo1279 horizontal resolution, corresponding to a spectral truncation at total wavenumber 1279 and an average grid-spacing of 9 km. Time integration is performed with a time step of \Delta t =450s. Vertically, the IFS uses a pressure-based hybrid \eta -coordinate with 137 model levels extending from the surface up to 0.01 hPa. The vertical discretization uses a third-order finite element method. 2.2 AIFS-Single-ML ECMWF has developed its own machine-learned global weather forecasting system. Here, AIFS-Single refers to the version of AIFS that is trained via a MSE based loss function. AIFS-Single has been running operationally since February 2024. It uses an encoder-processor-decoder architecture where the encoder maps the input state to a lower-resolution latent space using a Graph Neural Network (GNN) and a Transformer-based processor then forecasts the evolution of this latent state 6 hours forward in time. The decoder