ÌÇÐÄÊÓƵ

With heat waves among Europe's deadliest climate hazards, a team of scientists led by CMCC has developed a prediction system capable of providing helpful information four to seven weeks before summer, which gives valuable time to improve preparedness.

Trained on data from centuries of climate analysis up to recent years, the machine learning system has demonstrated an increase in forecast efficiency by drastically reducing the computational resources required, making these techniques accessible to a broader number of researchers and institutions.

The study "Seasonal forecasting of European heat waves using a feature selection framework," in Communications Earth & Environment, demonstrates how machine learning (ML) and artificial intelligence (AI) techniques are revolutionizing climate science by enabling more accurate, cost-effective predictions than traditional approaches.

Furthermore, where conventional dynamical forecasting systems require massive computational resources and struggle with reliability in northern European regions, this data-driven approach offers an alternative.

"ML will become a fundamental part of how we study climate variability," says McAdam. "This study has demonstrated the usefulness of ML in extreme event prediction, but it is only a first step in defining how we do that to receive interpretable and physically-meaningful results."

Heat waves cause devastating impacts across Europe, including agricultural losses, energy usage spikes, health crises, and increased mortality.

Recent deadly events in 2003, 2010, and 2022 underscore the urgent need for early warning systems that can help dampen the impacts of heat waves. This is particularly important as climate projections suggest further intensification of heat waves in the coming decades, making accurate seasonal forecasting crucial for saving lives.

"Early warning of extremely hot summers could help society prepare to mitigate against economic losses and reduce risk to life," McAdam explains. "Seasonal forecasts made in spring can, in principle, state whether a summer will be warmer than average."

Innovative methodology

The system employs an optimization-based feature selection framework that identifies the optimal combination of atmospheric, oceanic, and land variables to predict heat wave likelihood across Europe. Using ML techniques, the approach analyzes roughly 2,000 potential predictors to select the most predictive combinations for each geographic location.

The method not only matches, and in some cases outperforms, traditional forecast systems, but also provides information on which predictors were used in the process—a valuable scientific resource. The ability to pinpoint which atmospheric and oceanic predictors contribute most to forecast skill at different times and locations across Europe can inform future research into the physical mechanisms behind extreme heat events, for example.

For example, the research reveals that European soil moisture, temperature patterns and atmospheric circulation emerge as the most critical local predictors, while distant signals from the tropical Pacific and Atlantic also contribute to forecast skill.

A persistent challenge in seasonal forecasting has been poor performance over Scandinavia and northern-central Europe. In contrast, the new data-driven approach developed in the paper improves skill in these previously problematic areas.

One of the study's most innovative aspects involved training the ML system on paleoclimate simulations spanning years 0–1850, providing vastly more training data than available in observational records. Despite this unusual approach, the system successfully transferred its learning to accurately predict real-world heat waves from 1993–2016.

"There is not yet enough real-world data to train the forecast sufficiently, so the ML models actually learned about heat wave drivers in a model world but successfully applied the training to the real world," says McAdam.

A matter of efficiency

Not only increased efficiency but also a dramatic reduction in computational requirements make seasonal forecasting using this technique accessible to a broader range of researchers and institutions. While traditional dynamical systems require enormous supercomputing resources to run, this approach focuses specifically on heat wave prediction with minimal computational overhead.

"Our research has successfully extended data-driven ML-based forecasting to the seasonal timescale using a tiny fraction of the computational resources of traditional approaches," McAdam notes.

By providing reliable seasonal forecasts of extreme heat months in advance, the system enables proactive measures to reduce heat wave impacts on society and the economy.

This opens new possibilities for climate services across sectors, including agriculture, public health, energy management and emergency planning, as well as creating an opportunity to combine ML approaches with the dynamical system produced by CMCC and therefore leverage the strengths of both approaches.

The framework also has the potential to be adapted for other extreme events, start dates and target seasons, representing a significant milestone in CMCC's mission to advance climate science through innovative methodologies, and establish new standards for seasonal forecasting and climate risk assessment.