Mitigating Uncertainties in <scp>CMIP6</scp> Sea‐Level Simulations From Internal Variability Using the Semi‐Empirical Bayesian Ensemble

ABSTRACT Global climate models (GCMs) are essential to projecting sea‐level variations (SLV), yet at regional scales, their simulations frequently exhibit substantial biases associated with internal variability. Reducing these uncertainties in an interpretable way is critical for enhancing the reliability of regional projections. Here, we develop a Bayesian Model Averaging (BMA) ensemble strategy. GCM‐simulated SLV is used to construct a controlled baseline ensemble, against which additional physical signals are systematically integrated to quantify their contributions and assess improvements in projection accuracy at 1° resolutions between 60°S and 60°N. We evaluate contributions from 11 GCMs together with three observational datasets representing El Niño‐Southern Oscillation (ENSO), Pacific Decadal Oscillation (PDO), and Eddy Kinetic Energy (EKE). GCM‐simulated ENSO and PDO indices are also examined to test whether projections can be self‐corrected through internally generated variability, thereby offering potential benefits for future predictions without reliance on observations. The results show that the global mean RMSE of the 11‐model BMA ensemble is approximately 0.0692 m, with an overall improvement of 5.07% attributable to the inclusion of observed ENSO, PDO, and EKE data. These improvements are substantially more pronounced in localised regions. In both the in‐sample evaluation and the out‐of‐sample temporal cross‐validation, the error reductions display highly robust spatial structures, featuring ENSO‐like east–west reversals, PDO‐like horseshoe patterns, and significant improvements in eddy‐rich regions like the Gulf Stream and Kuroshio Extension. Although the CMIP6‐simulated ENSO and PDO signals are less effective than their observational counterparts, they nonetheless demonstrate potential corrective value, particularly in the case of the GCM‐simulated PDO.