Ocean MDTs are all in unformatted, real*4 files. Units in all cases are metres. Missing data and land are set to NaN. Native grid data means on the grid used by the model. This grids are given in the folder: model_grids. The dimensions nx,ny of the 2D fields are given in the file names (so file lengths should be 4*nx*ny). Latitude and longitude of the centre of each model grid cell are given. Regridded data are all on the same 1/8 degree grid which is 2880 longitudes by 1440 latitudes: longitudes (0.5,1.5,2.5 .... 2879.5)/8 degrees latitudes -90 + (0.5,1.5,2.5 .... 1439.5)/8 degrees 5-year means are 2D fields for each model, averaged over the 5 years 2003 to 2007 inclusive. Annual means are in files containing 3D datasets of size nx*ny*nt where nt is the number of years (file names say which years). e.g. if the filename includes '1993-2012' then nt=20. The native grid data have all been adjusted to have zero spatial mean. Some models do not have an Arctic, so the spatial means are not all directly comparable. The regridded data have been further adjusted so that the 5-year mean ecco2 model has zero spatial mean, and all other models have zero spatial mean difference from that reference model over their common domain. This makes the models more directly comparable. The spatial mean offsets applied are available in .txt files. The regridding has been done as follows. 1) Model coastal points have been identified and everywhere else set to missing 2) The strip of coastal points has been expanded by 12 iterations of replacing missing points with neighbouring valid points, taking care to reset to missing any points which correspond to land on the 1/8 degree grid (this avoids smoothing across land barriers). 3) Both the coastal strip and the complete model dataset are bilinearly interpolated onto the 1/8 degree grid. The interpolated complete model is then used as the base for what follows 4) Points which are coastal on the 1/8 degree grid, and ocean points which are missing, are replaced by values from the coastal strip. This ensures that all coastal points on the 1/8 degree grid come from model points which were coastal, if the model had land at all in that region. Step 3 involves slight extrapolation at the boundaries. In order to retain the coastal values, this is done as follows: Two fields are created: The field to be interpolated and a mask, initially set to 1 everywhere. Missing data in the field is set to zero, and the corresponding mask also set to zero. Both field and mask are interpolated (using zeros where needed) The field values are then divided by the mask values, effectively reweighting the zeros in the interpolation so they are not used. The final model datasets all have very similar patterns of land, but there can be differences where the model has no data (e.g. Arctic, Red Sea in some models), so the coastal strip does not fill the missing points. There may also be differences in the case of land barriers in some models (nemo12b in particular has some islands which do not appear in the 1/8 degree mask).