The observed warming of global mean surface temperature has been used to reduce uncertainty in future climate change and impact projections, but the information embedded in the spatial pattern of warming remains largely untapped. Here, we use machine learning to uncover spatially resolved emergent constraint relationships between 1971-2020 warming trends at individual grid cells and future global mean warming in a large collection of climate model simulations. This approach identifies key tropical and polar regions, with clear physical interpretations, where historical warming most effectively constrains future outcomes. Including the observed pattern information reduces the error variance in future global warming projections by about 70%, while commonly used constraints based solely on the global mean trend reduce that uncertainty by about 48%. These refined projections imply a higher likelihood of early exceedance of Paris Agreement thresholds. For example, the likelihood of exceeding 2 °C by mid-century under the SSP3-7.0 scenario is approximately 80%, compared to 70% when constrained by observed global mean trend alone. Our study highlights the potential of machine learning to uncover physically interpretable emergent constraints that improve future climate projections. This study presents a machine learning-based emergent constraint that reduces uncertainty in future warming projections by ~70% by identifying land and ocean areas that are the most informative of future warming. Earlier than anticipated exceedance of 2 °C warming becomes more likely with the better constrained projections.
Li et al. (Tue,) studied this question.