Sequential learning of climate change via a physical-parameter-based state-space model and Bayesian inference

Flexible decision-making strategies provide an alternative option for climate adaptation by considering future learning of climate change. A physical-parameter-based state-space model (SSM) with Bayesian inference is developed in this work to investigate reduction of uncertainty from more observations and facilitate flexible adaptation strategies. This SSM method integrates a two-layer, energy-balance model to describe global mean temperature response, models multiple sources of uncertainty such as climate sensitivity and aerosol forcing, and uses the informative priors from processing Global Climate Model simulations. Focusing on global mean temperature anomaly, which has important implications on policies and related impacts, the SSM is assessed by applying it to both historical and pseudo-observations (i.e., model simulations used as observations), assessing the posterior probabilities of physical parameters, and evaluating reduction of projection uncertainty. Some limitations of the method are observed, such as the sensitivity related to the adopted forcing time series. Comparing the end-of-the-century projections of global mean temperature sequentially made at year 2020, 2050, and 2080 using pseudo-observations, the reduction of uncertainty from the SSM is evident: the range of 95% prediction intervals on average decreases from 1.9°C in 2020 to 1.0°C in 2050, and to 0.6°C in 2080 under the Shared Socioeconomic Pathway (SSP) 2–4.5 (or from 2.7°C, to 1.2°C and to 0.7°C under SSP5-8.5). These results illustrate how the SSM framework provides probabilistic projections of climate change that can be sequentially updated with more observations, and this process can facilitate flexible adaptation strategies.