Assessment of Red Sea temperatures in CMIP5 models for present and future climate

each season) were then averaged to obtain the averaged seasonal correlation coefficient and RMSE. Seasonal cycle of the sea surface temperature. The ability of the models to capture regional differences was assessed through their capacity to reproduce the spatial pattern of the seasonal anomalies. The focus is on sea surface temperature, where the largest regional differ- ences are expected. In particular we compared modelled and observed seasonal anomalies in terms of spatial correlation and RMSE. The four values obtained (i.e., one for each season) were then averaged to obtain the averaged seasonal correlation coefficient and RMSE. Interannual variability. In order to assess the model skills in reproducing interannual variability, we computed the annual mean of the basin-average temperature field at each verti- cal level. We then calculated the standard deviation (std) to obtain a vertical profile of interan- nual std, which quantifies the amplitude of interannual signals at different depths. The shape and the values of the profiles were compared with observed ones in terms of correlation and RMSE. We further assessed the accuracy of the spatial patterns of sea surface temperature interannual std, again comparing them with the observed patterns in terms of spatial correla- tion and RMSE. Trends. An additional diagnostic that helps to identify potential problems with heat uptake in the models is the evaluation of trends in the recent decades. We cannot expect that GCM free runs reproduce realized decadal trends, so the goal of this diagnostic is to identify

models that are drifting, so reproducing unrealistic trends (i.e. trends exceeding what could be expected by natural variability or climate change). In this case, we did not include the GLORYS and KAUST models in the assessment because they cover a different time window, and the trend computation is very sensitive to the length of the record used for trend estimates. The slope of the linear trend of the basin averaged SST (τ) (i.e. the slope of a line fitted to the annual mean sea surface temperature), and its associated standard error (ε) were computed for each model and compared with those of TEMPERSEA. The difference in the trends between obser- vations and model outputs (θ = |τmod−τobs|) was compared with the square root of the qua- dratic sum of the standard errors of the fitting ðε ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ε2 mod þ ε2 obsÞ p . This has then be used to select suitable models (see section 3). Model skill assessment Area and volume In the Red Sea, 32 of 43 models were in the range of ±>25% with respect to the actual area, and 24 of 43 with respect to the actual volume (Fig 2). In the Outer Region, 41 of 43 and 39 of 43 meet that condition, for area and volume, respectively. As a reference, the global model GLORYS reaches a 91% (90.2%) of the actual area (volume), while the regional high-resolution KAUST model represents a 100.7% (99.7%) of the actual area (volume). Three of the CMIP5 GCMs did not include any representation of the Red Sea, and 9 of those including the Red Sea

represent it as a much smaller basin. Other 9 models show a reasonable surface area but were biased toward either a much shallower or deeper Red Sea, and were, accordingly, discarded. Conversely, in the Outer Region only 5 of those discarded models did not fulfil the require- ments (see S1 Table). Seasonally averaged vertical temperature profile In the Red Sea, the observed basin averaged temperature profile shows an anomaly with respect to the 3D basin average of ~2˚C at the surface during winter and spring (Fig 3, black line). In summer and autumn, the temperature anomaly exceeds 5˚C. Below 100 m the PLOS ONE Assessment of Red Sea temperatures in CMIP5 models for present and future climate PLOS ONE | https://doi.org/10.1371/journal.pone.0255505 July 30, 2021 6 / 26