The Euro-Mediterranean Center on Climate Change (CMCC) decadal prediction system

Abstract

Abstract. Decadal climate predictions, obtained by constraining the initial condition of a dynamical model through a truthful estimate of the observed climate state, provide an accurate assessment of near-term climate change and are a useful tool to inform decision-makers on future climate-related risks. Here we present results from the CMIP6 (Coupled Model Intercomparison Project Phase 6) Decadal Climate Prediction Project (DCPP) decadal hindcasts produced with the operational CMCC (Euro-Mediterranean Center on Climate Change) decadal prediction system (DPS), based on the fully coupled CMCC-CM2-SR5 dynamical model. A 20-member suite of 10-year retrospective forecasts, initialized every year from 1960 to 2020, is performed using a full-field initialization strategy. The predictive skill for key variables is assessed and compared with the skill of an ensemble of non-initialized historical simulations so as to quantify the added value of the initialization. In particular, the CMCC DPS is able to skillfully reproduce past climate surface and subsurface temperature fluctuations over large parts of the globe. The North Atlantic Ocean is the region that benefits the most from initialization, with the largest skill enhancement occurring over the subpolar region compared to historical simulations. On the other hand, the predictive skill over the Pacific Ocean rapidly decays with forecast time, especially over the North Pacific. In terms of precipitation, the skill of the CMCC DPS is significantly higher than that of the historical simulations over a few specific regions, including the Sahel, northern Eurasia, and over western and central Europe. The Atlantic multidecadal variability is also skillfully predicted, and this likely contributes to the skill found over remote areas through downstream influence, circulation changes, and teleconnections. Considering the relatively small ensemble size, a remarkable prediction skill is also found for the North Atlantic Oscillation, with maximum correlations obtained in the 1–9 lead year range. Systematic errors also affect the forecast quality of the CMCC DPS, featuring a prominent cold bias over the Northern Hemisphere, which is not found in the historical runs, suggesting that, in some areas, the adopted full-field initialization strategy likely perturbs the equilibrium state of the model climate quite significantly. The encouraging results obtained in this study indicate that climate variability over land can be predictable over a multiyear range, and they demonstrate that the CMCC DPS is a valuable addition to the current generation of DPSs. This stresses the need to further explore the potential of the near-term predictions, further improving future decadal systems and initialization methods, with the aim to provide a reliable tool to inform decision-makers on how regional climate will evolve in the next decade.

Gross primary productivity and the predictability of CO2: more uncertainty in what we predict than how well we predict it

“…The distinguishing element of CMCC-CM is the ocean, which is simulated by NEMO3.6, while sea ice is modelled by CICE4. The 10-member hindcast simulations are initialized every November (Nicolì et al, 2023). The ocean initial conditions are from CHOR (Yang et al, 2017) until 2010 and from CGLORSv7 (Storto and Masina, 2016) afterwards.…”

Deep Learning‐Based Seasonal Forecast of Sea Ice Considering Atmospheric Conditions

“…Dynamic models are typical methods constructed upon the physical linkages between ice and other climate factors in the Earth system. Common seasonal dynamic models include the Euro‐Mediterranean Center for Climate Change ‐ Seasonal Prediction System (Nicol I et al., 2023), ECMWF's seasonal forecasting system SEAS5 (Johnson et al., 2019), METEO‐FRANCE seasonal forecasting system, Global Seasonal Forecast System Version 5 (GloSea5) (MacLachlan et al., 2015), and so on. These seasonal dynamic models can provide operational forecasts for sea ice, but the requirements for extensive parameter initialization and computational resources limit their lightweight deployment and operation.…”