Comparative Evaluation of Decadal Predictions of Global <scp>SST</scp> Between <scp>CMIP5</scp> and <scp>CMIP6</scp> Datasets

ABSTRACT Accurate predictions of the Earth's near‐term (1 year and 2–5 year lead‐time) climate are crucial for informed decision‐making in various sectors such as agriculture, energy, public health, and infrastructure planning. Using the yearly initialized decadal hindcasts of the sea surface temperature (SST) from the CMIP5 and CMIP6 datasets, we evaluated their prediction skills over the North Pacific, North Atlantic, Indian Ocean, and tropical eastern Pacific for the next 5 years. In terms of spatial patterns, only the CMIP5 CanCM4 model exhibits a Pacific decadal oscillation (PDO) pattern that aligns with the observations. Although CMIP6 models cannot precisely replicate the spatial pattern of the PDO, their accurate prediction of North Pacific mid‐latitude SST indicates reliable regional forecasting. In the North Atlantic, all the models except the CMIP6 CanESM5 can reproduce a spatial pattern for the Atlantic multi‐decadal oscillation closely resembling observations. Evaluation of prediction skill over the Indian Ocean and tropical eastern Pacific was performed on an interannual scale, focusing on predictions at the first forecast lead year in terms of seasonal phase locking and prediction accuracy. Notably, the CMIP6 dataset exhibited superior performance compared to CMIP5 in the Indian Ocean. Across seasonality, all models effectively captured the seasonal peak of the Indian Ocean Dipole (IOD), occurring in September–November. CMIP6 demonstrated superior performance to CMIP5 in predicting IOD intensity, RMSE, and correlation coefficients. Evaluation over the tropical eastern Pacific revealed no significant improvement in prediction skill from CMIP5 to CMIP6. The heightened prediction skill for the IOD in the CMIP6 dataset, relative to the CMIP5 dataset, is primarily evident in the eastern tropical Indian Ocean. Additionally, models in CMIP6 could simulate a robust correlation between the dipole mode index and Niño 3.4 index, whereas those in CMIP5 could not, underscoring an advancement in predictive capabilities.