The passing of over three decades since tropical cyclone Tracy devastated Darwin on Christmas Day 1974 provided a reason to reflect upon the progress of numerical weather prediction over that period. Each decade was characterised by major model changes. In Australia, the 1970s saw the operational implementation of a 7-level baroclinic filtered model at a grid spacing of 254 km, with rudimentary representation of physical processes. The 1980s witnessed the operational introduction of a 12-level primitive equations model with a grid spacing of 150 km and enhanced representation of both the data assimilation approach and the model physics. Finally, in the 1990s the modeling capacity had reached the point where a 40-level model with a grid spacing of 5 km, such as the one employed here, could be run in near real-time, nested over the area of interest. Some research versions also had became available at even higher resolutions. Further improvements in the representation of physical processes and advanced data assimilation procedures, such as 4D-Var, were also available. Here, three models similar to those cited above were run out to 48 hours for the tropical cyclone Tracy case. In all models a bogussed vortex was required to provide an initial tropical cyclone circulation. However, the baroclinic model quickly lost tropical cyclone Tracy's circulation and it decayed so rapidly such that neither a circulation nor a track was discernable at 48 hours. The 1980s model was more sophisticated but still provided little gain over the 1970s model because its grid spacing remained well below that required for effective tropical cyclone modelling, especially for a storm as small as tropical cyclone Tracy. The 1990s model, which is much closer to those used operationally at present, in terms of data assimilation, resolution and physics, performed far better and provided a track for tropical cyclone Tracy that if available in 1974 could have alerted forecasters to a possible impact on Darwin. The results confirmed that major advances have taken place in tropical cyclone track forecasting, but to a lesser extent in predicting tropical cyclone intensity ( especially for very small tropical cyclones like Tracy) over the past 30 years or so. They also confirmed that at present, a gap remains between research and operations. Since the development of the 1990s model, further advances in data coverage, assimilation, model formulation and resolution have produced a new generation of 2000s models. These models have advanced cloud microphysics and data assimilation schemes and can be run in research mode at resolutions well below the 1 km required to provide more realistic predictions of small storms like tropical cyclone Tracy. The computational capacity soon will exist to run these sophisticated, very high resolution models in real-time.
