Advancing Subseasonal PredIctions at Reduced computational Effort

Informations projet

ASPIRE

N° de convention de subvention: 101077260

DOI

10.3030/101077260

Date de signature de la CE 12 Decembre 2022

Date de début 1 Septembre 2023

Date de fin 31 Août 2028

Financé au titre de

• European Research Council (ERC)

Coût total

€ 1 496 246,00

Contribution de l’UE

€ 1 496 246,00

Coordonné par

UNIVERSITAT ZU KOLN
Germany

Periodic Reporting for period 1 - ASPIRE (Advancing Subseasonal PredIctions at Reduced computational Effort)

Période du rapport: 2023-09-01 au 2026-02-28

Résumé du contexte et des objectifs généraux du projet

Extreme weather events, which are expected to become more frequent in a warming climate, pose a serious threat to humanity globally. Munich Re, the world’s largest reinsurance company, registered 980 natural catastrophes worldwide in 2020 alone. The five deadliest of these were weather-related. Flash floods and tropical cyclones resulted in more than 1,000 fatalities. In Europe, weather-related extremes caused economic losses averaging 12.5 billion Euros per year over the past decade and the 2003 European heatwave alone led to 70,000 excess deaths. Accurate predictions of weather extremes by the national weather services are indispensable so that preventive actions can be taken by the authorities and the public at an early stage. In particular, pro-active disaster mitigation efforts such as the provisioning of relief goods may take several weeks to implement. Reliable forecasts of the possible occurrence of weather extremes two weeks to two months ahead are therefore of immense value. In addition to such forecasts, reliable predictions on this so-called subseasonal time scale have a considerable socio-economic value for various other sectors. Examples are health, management decisions in agriculture and food security, or efficient planning in the energy sector with its growing share of renewable electricity generation and long-term energy storage systems.

The fundamental challenge of subseasonal forecasts is that, unlike ordinary medium-range weather forecasts that cover at most two weeks, the predictability cannot be gained solely from information on the initial atmospheric state. Due to the chaotic nature of the atmosphere, much of the memory of the atmospheric initial state is lost at this time scale. Further, the subseasonal time scale is yet too short for the slowly changing components of the Earth system, such as the ocean, to contribute significantly to predictability. One approach to account for the associated inherent unpredictability is to perform so-called ensemble forecasts which consider uncertainties in the initial state of the forecast and in the model formulation. These ensemble forecasts, however, become computationally costly especially when being run at a resolution high enough to accurately represent small-scale atmospheric processes. Leveraging recent progress in numerical modelling, machine learning and the availability of large sets of subseasonal weather forecasts, the project aims to reduce the computational costs while at the same time enhance the forecast skill on this time range.If successful, this approach would be a breakthrough towards improved operational weather forecasts at substantially lower computational costs, for a global socio-economic benefit.

Travail effectué depuis le début du projet jusqu’à la fin de la période considérée dans le rapport et principaux résultats atteints jusqu’à présent

Since the start of ASPIRE, the project team has pursued a two-pronged approach: (1) developing and refining modelling and machine-learning architectures to deliver efficient subseasonal forecasting, and (2) deploying diagnostic experiments to deepen our understanding of predictability and error sources at the subseasonal time-scale. In the following, we summarize the major scientific achievements to date, with a focus on the three most important publication outcomes.
1) Model architecture design, training methodology and computational efficiency
One of the early major achievements is the publication of “Architectural insights into and training methodology optimisation of Pangu-Weather” (To et al., 2024). In this collaborative effort with colleagues from the computer science department at the Karlsruhe Institute of Technology, we performed an in-depth ablation study of the state-of-the-art data-driven weather forecasting model Pangu‑Weather (PGW), focusing on architectural choices (e.g. learned positional embeddings, relative bias, 2-D attention) and training schedule optimisation (e.g. variable-specific weighted cosine loss scheduling) to improve both performance and computational cost. Key technical achievements from this activity demonstrate that certain ablated variants of the PGW architecture achieve almost equivalent validation loss with significantly lower parameter-count/training cost. By quantifying compute requirements for different model variants, we identified a way to reduce computational costs by up to 20-30%. This achievement addresses directly the ASPIRE objective of developing computationally efficient modelling strategies that maintain or enhance forecast skill, thereby enabling larger-ensemble experiments and longer hindcasts.

2) Uncertainty quantification for data-driven weather models
In another major activity the team advanced methods for uncertainty quantification (UQ) in data-driven forecasting frameworks. This is captured in the paper “Uncertainty quantification for data-driven weather models” (Bülte et al. 2026). In this study, we systematically contrast initial‐condition perturbation (ensemble) approaches with statistical/machine‐learning post-hoc UQ methods for the deterministic PGW model. We demonstrate that PGW + post-hoc UQ methods can yield probabilistic forecasts that outperform the physics‐based ensemble for lead times of up to ~5 days in the European region. This work advances both the scientific understanding of how to quantify uncertainty in ML‐forecast systems and the technical path towards operationally viable probabilistic subseasonal forecasts.

3) Diagnostic predictability experiments for subseasonal timescales
The third major achievement is the pre-print paper “Relaxation experiments in ML-based weather prediction models to study subseasonal predictability” (Li et al., 2026). In this study, we newly introduce nudging experiments in machine learning based weather prediction (MLWP) models to identify sources of forecast error. We compare the forecast skill in week 3-4 forecasts of two major precipitation events in western North America and contrast MLWP models with a traditional numerical weather-prediction model. We conclude that the relaxation experiments implemented in MLWP models provide a low-cost diagnostic to test influence of specific regions (e.g. tropics) on extratropical subseasonal prediction skill. Results show that the MLWP models already demonstrate baseline skill at subseasonal lead times in these case studies, and that imposing tropical relaxation improves the forecast skill. It thus contributes to ASPIRE’s objective of deepening scientific understanding of mechanisms governing subseasonal predictability.

Progrès au-delà de l’état des connaissances et impact potentiel prévu (y compris l’impact socio-économique et les conséquences sociétales plus larges du projet jusqu’à présent)

The expected impacts of ASPIRE extend beyond the immediate academic community. Improved and more efficient subseasonal forecasts can enhance preparedness in sectors such as energy, agriculture, water management, and disaster risk reduction. By lowering the computational barriers to high-quality forecasts, ASPIRE’s approaches make advanced prediction capabilities accessible to a wider range of institutions, including those with limited computing infrastructure. Further research is especially needed on how to ideally couple the forecasting systems developed here to applications. Initial steps to couple the models used by ASPIRE to energy system models have been taken. To maximise the long-term impact and uptake of ASPIRE’s results, several key enablers have been identified.First we need to continuously test our methods across different regions, seasons, and variables to demonstrate robustness and generalisability. Continued collaboration between atmospheric scientists and computer scientists will be critical to ensure the interpretability and physical consistency of data-driven approaches.