REE10: Modelling and predicting climate change

Background

Climate and weather simulations have been found to consistently under-represent modes of variability in the Earth-atmosphere system. This is, at least in part, due to bulk-formula parameterisations used to represent unresolved physical processes. Developing stochastic representations of unresolved processes that properly account for uncertainties in their behaviour will hopefully allow for more accurate and realistic probabilistic forecasting and simulations.

Techniques and Challenges

One proposed method is to construct a framework for obtaining stochastic sub-grid schemes directly from the simulated equations using stochastic calculus and a two-scale variational multiscale approach. In order to obtain sensible results, ensembles of forecasts are required, which increases computational cost. Inherently unpredictable computer hardware has been recently proposed as a way to improve power/performance ratios, which could benefit probabilistic forecasting by providing the random contributions. A third strand of the project is to investigate the predictability of nonlinear partial differential equations (PDEs) with turbulent characteristics when simulated at resolutions far from the convergence regime.

Results

The stochastic variational multiscale approach has been implemented on simple test problems with promising results. A predictability experiment using the surface quasi-geostrophic equations has begun, starting with a new numerical implementation. A ‘truth’ run at very high resolution has been carried out, and lower resolution ‘perturbed’ simulations are ongoing. Using a model of hardware faults provided by collaborators, we can now emulate numerical simulations running on ‘stochastic hardware’. A comparison between the impact of faults on implicit and explicit numerical schemes is ongoing.

The Future

Predictability studies with the surface quasi-geostrophic equation (and possibly others) should help identify what is needed for numerical simulations to be ‘statistically consistent’ over long periods of time. Stochastic approaches for weather and climate will allow more nuanced probabilistic predictions to be made, and the possibility of running these on ‘stochastic’ hardware will lead to power savings and performance gains. Another area of future work is the impact of including a description of the hysteretic soil-water behaviour in weather and climate models.

References

Palmer T.N.: Towards the Probabilistic Earth-System Simulator: A Vision for the Future of Climate and Weather Prediction, submitted to Q. Jour. Roy. Met. Soc., 2011

Shanbhag N.R., Abdallah R.A., Kumar R., Jones D.L.: Stochastic Computation, DAC '10 Proceedings of the 47th Design Automation Conference, 2010

This work is funded by the Oxford Martin School