Dry season fires in northern Australia each year cause an abrupt change in the savannah vegetation properties. These fires may occur naturally, accidentally, or as part of a land management strategy. The fire regime of northern Australia is distinctly seasonal (Tapper et al. 1993), with early dry season (May through June) fires being generally of low intensity, and late dry season and pre-monsoonal (August to October) fires up to an order of magnitude greater intensity. Over the last 50 years, there has been a strong shift in the fire regime in northern Australia, from a regime characterised by early dry season burning, to more destructive late dry season fires (Williams et al. 1997, Williams et al. 2001). This not only influences local boundary layer processes but there is also some evidence which suggests an influence on the Australian Monsoon and especially the associated rainfall patterns large regions of Australia heavily rely on (Beringer et al. 2003).

We use the CSIRO Conformal-Cubic Atmospheric Model (C-CAM) with a stretched grid, so that the model has a resolution of about 60 km over Australia, and far field nudging provided by the NCEP/NCAR Reanalysis. In the first phase of the study, changes of the vegetation properties by fires are implemented in the C-CAM land surface model by modifying key vegetation parameters such as albedo, leaf area index and roughness length, accounting for the vegetation succession after the fires, based on observational data and plant physiological information. In a sensitivity study, results from a 21-year model run (1979 to 1999) with 5-years spin up time (1974 to 1978) are compared to simulations without any fires. Observational data (see Tapper et al., this session) are used to assess the suitability of the approach and the model. The impact on the atmosphere is analysed by response metrics for each year, i.e. Australian Monsoon Index, location and central pressure of Cloncurry and Pilbara heat lows, ITCZ properties over Australia and precipitation parameters for special areas.

In a second phase, the forcing perturbation sets (timing of fires, intensity, area burned, length of succession period) are varied continuously within identified ranges using relatively broad intervals and the Latin hypercube sampling technique (Lynch et al. 2001). To optimize performance and efficiency of the experiment, the software tool NIMROD/G (Buyya et al. 2000) is used to control and perform the experiments in parallel on several Linux clusters using a grid computing paradigm. The integrative metrics of multiple realisations of the physical model are used together with the multidimensional forcing variable space to express the response space by a reduced form nonlinear statistical model. It can be used to explore the physical relationships between the fires’ effects and the impact on the Australian Monsoon in a computationally efficient way. Based on these results a further optimisation is reached by adapting the search heuristics of the NIMROD/O optimisation tool (Abramson et al. 2000) which is then able to search the forcing space based on nonlinearities and thresholds in the response space.

At the present stage we present results from phase one and some preliminary work from phase two.