Evaluation of global ocean–sea-ice model simulations based on the experimental protocols of the Ocean Model Intercomparison Project phase 2 (OMIP-2)

Abstract. We present a new framework for global ocean–sea-ice model simulations based on phase 2 of the Ocean Model Intercomparison Project (OMIP-2), making use of the JRA55-do atmospheric dataset. We motivate the use of OMIP-2 over the framework for the first phase of OMIP (OMIP-1), previously referred to as the Coordinated Ocean–ice Reference Experiments (CORE), via the evaluation of OMIP-1 and OMIP-2 simulations from eleven (11) state-of-the-science global ocean–sea-ice models. In the present evaluation, multi-model means are calculated separately for the OMIP-1 and OMIP-2 simulations and overall performances are assessed considering metrics commonly used by ocean modelers. Many features are very similar between OMIP-1 and OMIP-2 simulations, and yet we also identify key improvements in transitioning from OMIP-1 to OMIP-2. For example, the sea surface temperature of the OMIP-2 simulations reproduce the observed global warming during the 1980s and 1990s, as well as the warming hiatus in the 2000s and the more recent accelerated warming, which were absent in OMIP-1, noting that OMIP-1 forcing stopped in 2009. A negative bias in the sea-ice concentration in summer of both hemispheres in OMIP-1 is significantly reduced in OMIP-2. The overall reproducibility of both seasonal and interannual variations in sea surface temperature and sea surface height (dynamic sea level) is improved in OMIP-2. Many of the remaining common model biases may be attributed either to errors in representing important processes in ocean–sea-ice models, some of which are expected to be reduced by using finer horizontal and/or vertical resolutions, or to shared biases in the atmospheric forcing. In particular, further efforts are warranted to reduce remaining biases in OMIP-2 such as those related to the erroneous representation of deep and bottom water formations and circulations. We suggest that such problems can be resolved through collaboration between those developing models (including parameterizations) and forcing datasets.