Deep Learning Meets Teleconnections: Improving S2S Predictions for European Winter Weather Journal: Machine Learning: Earth
Philine L. Bommer, Marlene Kretschmer, Fiona R. Spuler, Kirill Bykov, Marina M.-C. Höhne · 2025
Read Deep Learning Meets Teleconnections: Improving S2S Predictions for European Winter Weather Journal: Machine Learning: Earth by Philine L. Bommer, Marlene Kretschmer, Fiona R. Spuler, Kirill Bykov, Marina M.-C. Höhne (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-2504-07625.txt.
Introduction
Philine Bommer and co-authors test whether deep learning can improve forecasts of European winter weather several weeks ahead. Such subseasonal forecasts aid early warning but fight against the chaos of the climate system. Distant drivers, called teleconnections, open windows of better skill: the stratospheric polar vortex and the Madden-Julian Oscillation are two of them. The team asks how much these drivers help a learning model.
Three models predict North Atlantic-European weather regimes six weeks out. One uses past regimes alone, one adds the two driver indices, and one reads full stratospheric wind and tropical radiation fields through a vision transformer. Adding driver data lifts skill at longer ranges. The field-based model beats the operational hindcasts of a major weather center beyond week four, chiefly on two regime types. The authors also link high-confidence forecasts to known driver patterns, and spot some new ones.
Catalog record
| CREATOR | Philine L. Bommer, Marlene Kretschmer, Fiona R. Spuler, Kirill Bykov, Marina M.-C. Höhne |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-07625
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Deep Learning Meets Teleconnections: Improving S2S Predictions for European Winter Weather Journal: Machine Learning: Earth Philine L. Bommer, Marlene Kretschmer, Fiona R. Spuler, Kirill Bykov, Marina M.-C. Höhne Abstract Predictions on subseasonal-to-seasonal ( S2S ) timescales—ranging from two weeks to two months—are crucial for early warning systems but remain challenging owing to chaos in the climate system. Teleconnections, such as the stratospheric polar vortex (SPV) and Madden-Julian Oscillation (MJO), offer windows of enhanced predictability, however, their complex interactions remain underutilized in operational forecasting. Here, we developed and evaluated deep learning architectures to predict North Atlantic-European (NAE) weather regimes, systematically assessing the role of remote drivers in improving S2S forecast skill of deep learning models. We implemented (1) a Long Short-term Memory (LSTM) network predicting the NAE regimes of the next six weeks based on previous regimes, (2) an Index-LSTM incorporating SPV and MJO indices, and (3) a ViT-LSTM using a Vision Transformer to directly encode stratospheric wind and tropical outgoing longwave radiation fields. These models are compared with operational hindcasts as well as other AI models. Our results show that leveraging teleconnection information enhances skill at longer lead times. Notably, the ViT-LSTM outperforms ECMWF’s subseasonal hindcasts beyond week 4 by improving Scandinavian Blocking (SB) and Atlantic Ridge (AR) predictions. Analysis of high-confidence predictions reveals that NAO − - , SB, and AR opportunity forecasts can be associated with SPV variability and MJO phase patterns aligning with established pathways, also indicating new patterns. Overall, our work demonstrates that encoding physically meaningful climate fields can enhance S2S prediction skill, advancing AI-driven subseasonal forecast. Moreover, the experiments highlight the potential of deep learning methods as investigative tools, providing new insights into atmospheric dynamics and predictability. Keywords : S2S, Deep Learning, Teleconnections, Forecasting, European weather regimes 1 Introduction Extreme weather events are becoming increasingly frequent and severe, posing a significant threat to societies, economies, and ecosystems. Improving longer lead-times predictions is essential for early warning systems and disaster mitigation, helping to reduce economic damage and humanitarian losses ( Coughlan et al. 2016 ) . However, such subseasonal-to-seasonal ( S2S ) forecasts—spanning two weeks to two months— remain strongly limited in skill due to the chaotic nature of the climate system ( Vitart et al. 2017 ; Vitart & Robertson 2018 ) . Predictability at S2S timescales arises due to teleconnections, where anomalies in one region can influence the persistence and transition of weather patterns in distant locations ( Vautard 1990 ; Seager et al. 2010 ; Nielsen et al. 2022 ) through wave propagation and large-scale circulation shifts ( Yamagami & Matsueda 2020 ) . In the North Atlantic–European (NAE) sector, the influence of teleconnections can be analyzed through the occurrence of weather regimes such as the positive (NAO+) and negative phase (NAO-) of the North Atlantic Oscillation, Atlantic Ridge (AR), and Scandinavian Blocking (SB). These persistent circulation patterns shape regional weather including extreme events ( Vautard 1990 ; Cattiaux et al. 2010 ; Seager et al. 2010 ; Luo et al. 2020 ; Ardilouze et al. 2021 ; Nielsen et al. 2022 ) . Among the key drivers of these regimes is the stratospheric polar vortex (SPV), a persistent cyclonic circulation in the Arctic stratosphere. Sudden stratospheric warming—rapid temperature increases in the polar stratosphere coinciding with a weakened flow—can disrupt the SPV, shifting the jet stream southward and triggering prolonged cold spells over northern Eurasia ( Kretschmer et al. 2018 ; Domeisen et al. 2020 ; Spaeth et al. 2024 ) . Similarly, the Madden-Julian Oscillation (MJO), an eastward-propagating tropical convective system, has been shown to modulate the NAE regimes, e.g. influencing the likelihood of negative (NAO - ) and positive North Atlantic Oscillation (NAO + ) phase up to 20 days later ( Cassou 2008 ; Lee et al. 2019 ; Lee et al. 2020 ; Robertson et al. 2020 ; Nardi et al. 2020 ) . The variability in these teleconnection drivers, thus, offers windows of enhanced predictability that can be leveraged to improve extreme weather forecasts ( Mariotti et al. 2020 ) . Despite recent research progress, our understanding of teleconnections remains limited, largely due to the complexity of their interactions ( Kretschmer et al. 2017 ) . In particular, it has been shown that most numerical weather models might underestimate established teleconnection pathways, e.g. struggling with ENSO variability as well as MJO phases influencing European weather or capturing stratosphere-troposphere couplings and the SPV ( Robertson et al. 2020 ; Spaeth et al. 2024a ; Rivière et al. 2024 ; Wu et al. 2024 ; Garfinkel et al. 2025 ) . Consequently, operational forecast systems struggle to accurately capture climate and weather dynamics beyond two weeks, limiting their utility for impact-driven decision-making and extreme weather preparedness. Machine learning approaches, particularly deep learning (DL), are showing promise not only in the weather domain but also in the S2S domain. While early efforts focused on convolutional recurrent models for spatiotemporal precipitation forecasting ( Shi et al. 2015 ) , later studies have already discussed the general potential of deep learning in S2S prediction ( Cohen et al. 2019 ) . Progress has continued with the introduction of generative neural weather models ( Weyn et al. 2021 ) and explainable AI (XAI) frameworks even identifying potential teleconnections and windows of enhanced predictability ( Mayer & Barnes 2020 ) . In recent years, works like Castro et al. 2021 , Peng et al. 2021 , and Mouatadid et al. 2022 have demonstrated the potential of hybrid data-driven models for teleconnection-aware prediction, physics-informed networks, and ML-driven bias correction of S2S forecasts. Advances include transformer-based architectures for extended-range forecasts ( Zhang et al. 2024 ) and multi-modal approaches combining physical and data-driven information ( Pérez-Carrasquilla & Molina 2024 ) , underscoring a growing trend toward more interpretable and skillful DL tools in S2S forecasting. Nonetheless, ML for S2S forecasting remains in its infancy, facing two fundamental challenges: first, capturing multi-timescale climate dynamics, and second, effectively representing teleconnections within ML architectures. To address these challenges and systematically evaluate the role of teleconnection drivers in S2S prediction skill, we develop DL approaches to forecast NAE regimes in boreal winter. We develop and compare three architectures with increasing complexity: 1. a basic Long Short-term Memory network (LSTM)—an LSTM trained to predict NAE regimes for six weeks, using only the past six weeks of regime states as input. 2. An index-augmented LSTM ( Index-LSTM )—an extension of LSTM by incorporating physical driver indices—the stratospheric polar vortex (SPV) strength and the Madden-Julian Oscillation (MJO) phase. 3. A spatiotemporal model ( ViT-LSTM )—an encoder-decoder model, consisting of a Vision Transformer (ViT) and an LSTM model that integrates the raw climate fields of zonal winds at 10 hPa (u10) over the polar region and outgoing longwave radiation (OLR) over the tropics—to allow the network to directly learn spatiotemporal teleconnection information. Using the well-established NAE regime framework, where MJO and SPV variability play a key role in predictability ( Spuler et al. 2025 ) , we demonstrate that physically meaningful drivers can enhance the S2S forecast skill. Moreover, we compare our predictions with other established machine learning approaches (Aurora-based architecture ( Bodnar et al. 2024 ) and logistic regression), as well as the dynamical