A fully prognostic Artic ice-ocean model is used to study the interannual variability of deepwater formation in the Greenland Sea gyre based on the simulations for the Arctic ice-ocean system for the period 1960--1985. The model uses monthly climatology for thermodynamic forcing components (such as air temperature and cloudiness), together with constant annual net precipitation and river runoff. The daily wind forcing is derived from analyzed sea level air pressures from the National Center for Atmospheric Research. In summary, the model shows that the occurrence of deep convection in the Greenland Sea gyre is controlled by the extensive Fram Strait ice export and/or local wind conditions in the Greenland Sea. In the latter case the weakening of local wind curl allows the Polar Front to move eastward. The movement of the Polar Front causes adverse ice conditions, often together with much larger than normal ice export from the Artic, such as in 1968, which can block convection in the gyre. The density difference between upper and lower layers is investigated as an indication of water mass formation through convection, occurring as strong diffusion in the model.

The model-simulated density difference between the average top 100 m and deep levels reveals that the period 1960--1985 had only a few distinct years with weak stratification, and, especially, the model predicts no deep convection since the mid-1970s. The common factor for the years of the weakest stratification is extreme wintertime wind events. Deep convection is also evident as a decrease of the model-predicted heat content of the upper 2000 m which can, to a high degree, be explained by local heat loss. The salt content changes in the Greenland gyre are independent of the heat content changes and are determined by the advection of fresh of salty anomalies, either from the Arctic or from the frontal movement. The model indicates that the surface salinity flux favors freshening, even in the wintertime, so that brine rejection is less important in buoyancy loss than heat loss in leading to deep convection. The model shows that the system has a memory; deep convection in one year will precondition the following year.