A novel initialization technique for decadal climate predictions

Model initialization is a matter of transferring the observed information available at the start of a forecast to the model. An optimal initialization is generally recognized to be able to improve climate predictions up to a few years ahead. However, systematic errors in models make the initializati...

Full description

Bibliographic Details
Published in:Frontiers in Climate
Main Authors: Volpi, Danila, Meccia, Virna L., Guemas, Virginie, Ortega Montilla, Pablo, Bilbao, Roberto, Doblas-Reyes, Francisco, Amaral, Arthur, Echevarria, Pablo, Mahmood, Rashed, Corti, Susanna
Other Authors: Barcelona Supercomputing Center
Format: Article in Journal/Newspaper
Language:English
Published: Frontiers Media 2021
Subjects:
Online Access:http://hdl.handle.net/2117/364823
https://doi.org/10.3389/fclim.2021.681127
Description
Summary:Model initialization is a matter of transferring the observed information available at the start of a forecast to the model. An optimal initialization is generally recognized to be able to improve climate predictions up to a few years ahead. However, systematic errors in models make the initialization process challenging. When the observed information is transferred to the model at the initialization time, the discrepancy between the observed and model mean climate causes the drift of the prediction toward the model-biased attractor. Although such drifts can be generally accounted for with a posteriori bias correction techniques, the bias evolving along the prediction might affect the variability that we aim at predicting, and disentangling the small magnitude of the climate signal from the initial drift to be removed represents a challenge. In this study, we present an innovative initialization technique that aims at reducing the initial drift by performing a quantile matching between the observed state at the initialization time and the model state distribution. The adjusted initial state belongs to the model attractor and the observed variability amplitude is scaled toward the model one. Multi-annual climate predictions integrated for 5 years and run with the EC-Earth3 Global Coupled Model have been initialized with this novel methodology, and their prediction skill has been compared with the non-initialized historical simulations from CMIP6 and with the same decadal prediction system but based on full-field initialization. We perform a skill assessment of the surface temperature, the heat content in the ocean upper layers, the sea level pressure, and the barotropic ocean circulation. The added value of the quantile matching initialization is shown in the North Atlantic subpolar region and over the North Pacific surface temperature as well as for the ocean heat content up to 5 years. Improvements are also found in the predictive skill of the Atlantic Meridional Overturning Circulation and the barotropic stream function in the Labrador Sea throughout the 5 forecast years when compared to the full field method. This study was supported by the project LISTEN funded by the European Commission Horizon 2020 Marie Skłodowska-Curie Actions - IF (GA 799930). The authors thankfully acknowledge the computer resources from the ECMWF special project INCIPIT (spitvolp) and the technical assistance provided by ECMWF and BSC. The climate simulations have been performed using Autosubmit workflow manager (Manubens-Gil et al., 2016). The authors thank Paolo Davini for providing the restart files of the historical simulation used to implement the quantile matching. We acknowledge Saskia Loosveldt and the earthdiags and ESMValTool suite developers, as well as the startR and s2dverification (Manubens et al., 2018) software packages developers, as these tools were used to postprocess, analyze, and visualize the results presented in this work. PO acknowledges support by the Spanish Ministry of Economy, Industry and Competitiveness through the Ramon y Cajal grant (RYC-2017-22772). Peer Reviewed Postprint (published version)