Weather Emulators Push the Frontier of Heat Extremes Forecasting

For more than half a century, meteorology has accepted a fundamental limit: weather cannot be predicted beyond two weeks. This boundary, rooted in the chaotic nature of the atmosphere, has shaped generations of forecasting science, defining the boundary between weather and climate forecasting and constraining our ability to anticipate high-impact extremes. At the same time, extreme heat has emerged as the deadliest climate-related hazard worldwide, underscoring the urgent need for reliable early warnings at lead times relevant for public health, energy systems, and disaster risk reduction. Recent advances in deep learning have produced a new class of global weather models accelerating progress in forecasting, raising the question of whether the traditional two-week limit is beginning to shift.

Here we evaluate six state-of-the-art deep learning weather emulators — Pangu-Weather, FuXi, ArchesWeather, AIFS, GraphCast and Aurora — alongside leading dynamical approaches and statistical baselines in forecasting global surface temperature and extreme heat events at a 14-day lead time. Models are evaluated using a suite of metrics, considering global temperature and extreme heat forecasting in both regression and classification settings. Several emulators rival or even surpass physics-based forecasts for temperature, but struggle to balance deterministic skill with realistic spectral properties. While all models display predictive skill for extreme heat, their predictions are deterministic and often inaccurate, offering little insight into uncertainty and limiting their reliability. Overall, results demonstrate that deep learning is starting to extend the frontiers of deterministic predictability. However, key limitations remain that constrain their applicability for operational early-warning systems, highlighting the need for reliable probabilistic approaches in a rapidly warming climate.