Benchmarking Deep Learning Models for Probabilistic Subseasonal Forecasting of Heat Extremes

Predicting climate extremes such as droughts, heatwaves, and heat stress episodes remains a critical challenge in Earth system sciences. Current state-of-the-art methods often fail to deliver reliable forecasts, especially at subseasonal-to-seasonal (S2S) timescales (i.e., from two weeks to two months in advance). As global climate variability continues evolving, the need for advanced, trustworthy, data-driven forecasting methodologies has never been more pressing.

Extended numerical weather prediction systems, such as those led by the European Centre for Medium-Range Weather Forecasts (ECMWF), remain the primary method for S2S prediction (Vitart & Robertson, 2018). While recent deep learning approaches have demonstrated remarkable competitive performance (e.g. Olivetti & Messori, 2024), proposed models predominantly focus on global-scale average weather predictions, overlooking critical local-scale extreme events (Pasche et al., 2024). Moreover, creating accurate probabilistic forecasts conditioned on the initial state remains a significant challenge within the scientific community. In the context of weather forecasting, traditional statistical methods, such as ensemble-based techniques that generate multiple forecasts to estimate uncertainty, are commonly used. These approaches include techniques such as introducing noise into initial states, varying neural network parameters, or training generative models. While generative models offer the most robust solutions, they demand substantial computational resources and extensive data availability.

Here, we evaluate several state-of-the-art dynamical weather forecasting systems, such as those of ECMWF and the National Centers for Environmental Prediction (NCEP), together with recently-proposed deep learning models on their ability to predict extreme heatwaves across all continents at S2S timescales. Since uncertainty quantification is essential for supporting practical decision-making, we focus on deep learning models that provide probabilistic forecasts and have publicly available source code. These include FourCastNet, proposed by Kurth et al. (2023), as well as ArchesWeather and ArchesWeatherGen, developed by Couairon et al. (2024). This analysis underscores the limitations of contemporary deep learning and dynamical weather forecasting systems in reliably and probabilistically predicting S2S extremes, while also providing a valuable benchmark to guide future research efforts.

 

References:

Couairon, G., Singh, R., Charantonis, A., Lessig, C., & Monteleoni, C. (2024). ArchesWeather & ArchesWeatherGen: a deterministic and generative model for efficient ML weather forecasting. arXiv preprint arXiv:2412.12971.

Kurth, T., Subramanian, S., Harrington, P., Pathak, J., Mardani, M., Hall, D., Miele, A., Kashinath, K., & Anandkumar, A. (2023). FourCastNet: Accelerating global high-resolution weather forecasting using adaptive Fourier neural operators. Proceedings of the Platform for Advanced Scientific Computing Conference (PASC '23), Article 13, 1–11. https://doi.org/10.1145/3592979.3593412

Olivetti, L., & Messori, G. (2024). Do data-driven models beat numerical models in forecasting weather extremes? A comparison of IFS HRES, Pangu-Weather, and GraphCast. Geoscientific Model Development, 17(21), 7915-7962.

Pasche, O. C., Wider, J., Zhang, Z., Zscheischler, J., & Engelke, S. (2025). Validating Deep Learning Weather Forecast Models on Recent High-Impact Extreme Events. Artificial Intelligence for the Earth Systems, 4(1), e240033. https://doi.org/10.1175/AIES-D-24-0033.1

Vitart, F., & Robertson, A. W. (2018). The sub-seasonal to seasonal prediction project (S2S) and the prediction of extreme events. npj climate and atmospheric science, 1(1), 3.