Teleconnection-driven predictability of the 2023 Brazilian heat wave

Understanding the predictability and humidity characteristics of South American heat extremes remains a key challenge as their frequency accelerates under global warming. Here, we combine temperature regionalization, air-mass typing, and a teleconnection-driven LSTM framework to investigate these questions, using Brazil's record-breaking November 2023 heat wave as a focal case. Area-mean temperatures during the event exceeded the 1940–2023 climatology by ~ 2°C, and a new national record of 44.8°C was reported. Using rotated principal component analysis, we identify three temperature-coherent regions across Brazil and show that the November 2023 event was exceptional in affecting all regions simultaneously. Composite analyses reveal that positive temperature extremes are consistently preceded by an intensified, westward-displaced South Atlantic Subtropical High, which suppresses convection and enhances surface heating through subsidence. The event also coincided with the lowest nationwide November soil moisture since 1964, consistent with land-atmosphere feedbacks that amplify extreme heat. Using the Gridded Weather Typing Classification (GWTC-2), we distinguish dry heat over the central interior from humid heat along the northeastern Atlantic coast, a distinction with direct implications for health risk assessment. Finally, we demonstrate that selected oceanic indices explain approximately 78% of the variability in temperature extremes within a Long Short-Term Memory neural network, revealing substantial teleconnection-driven predictability. These results highlight that Brazilian heat extremes arise from coupled land-atmosphere-ocean processes and that this predictability can support improved early warning systems.