Abstract
We train convolutional neural networks (CNNs) to predict the
distribution of peak warming given the current state of the climate
system and varying additional cumulative CO2 emissions. Even if the most
ambitious decarbonization goals are achieved, there is >99%
probability of exceeding 1.5˚C of long-term global warming,
approximately even odds of reaching 2˚C, and high likelihood that the
hottest year globally exceeds 2023 by at least 0.5˚C. Further, for more
gradual decarbonization that does not reach net-zero CO2 emissions this
century, there is >90% probability that the hottest annual
global temperature anomaly is twice as large as 2023. The fact that our
framework makes highly accurate out-of-sample predictions of the hottest
historical year provides confidence in the predicted future
probabilities. Given the non-linear sensitivity of many natural and
human systems, our results suggest substantial risks from the extreme
local conditions that could result from globally hot years during rapid
decarbonization.