Underestimation of the projected 21st century increase in drought duration, according to emergent constraint: consistent result in CMIP5 and CMIP6 models

Droughts are defined as one of the most devastating natural disasters of modern times and a key challenge faced under climate change. The complexity of interacting physical processes that underlie the shortage of rainfall in climate models hampers accurate representation of present-day droughts, and leads to differences in their responses to increased greenhouse gas (GHG) concentrations in the future. As a result, the confidence in drought projections is currently defined as ‘medium to low' by the Intergovernmental Panel on Climate Change (IPCC), and reducing this uncertainty remains one of the main goals in coming years, with significant benefits for human and natural systems. 

In this study we explore a relationship between biases in simulated present-day values of longest annual drought (LAD) and future projections of LAD in an ensemble of CMIP5 and CMIP6 models. We find that present-day model bias explains almost 95 % of the future uncertainty in LAD by the end of the 21st century, attributed to the well-known precipitation simulation errors: “drier” models with longer annual droughts at present tend to predict larger LAD values worldwide in the future, as well as a stronger response to GHG forcing in LAD, which is significant in more than 40 % of the global land area.

Substituting observational LAD estimates from satellite data into this model-revealed “present–future relationlarship” suggests that the 21st century global mean increase in duration of annual meteorological droughts could be significantly larger than predicted by the CMIP5 and CMIP6 model ensembles. This emergent constraint reduces global mean uncertainty range in future LAD estimates from 45–100 to 75–90 days, a level more typical of the prediction range of “drier” models. The findings reveal world regions where climate change may cause stronger meteorological drought aggravation than expected, and emphasise the importance of reducing model errors, which are presently largely owed to rain biases, to increase confidence in future predictions.