More accuracy with less precision: a new route to high resolution climate prediction

The climate system is a nonlinear multi-scale dynamical system. Reliable predictions of climate over the coming century are
critical to guide mitigation policy, to aid decisions about infrastructure to adapt to climate change, and to determine whether geogengineering
solutions are safe alternatives to zero-carbon emissions. It is clear that reliable predictions are not possible without accurate simulations of the
hydrological cycle, including cloud systems. However it will be many decades before computers are powerful enough to be able to simulate clouds directly from the
Navier-Stokes equations. Is there an alternative? He we discuss the role of imprecise computing as an emerging technology in climate simulation. The basic premise is that the closure or parametrisation process in weather and climate prediction is fundamentally stochastic in nature. This implies an irreducible level of imprecision in forecasting climate, represented by the use of forecast probability distributions as primitive prognostic outputs. We then discuss the application of techniques by which data is moved through a computer
at variable levels of precision, and assess the minimal level of bit precision which will not degrade solutions over and above this irreducible level of imprecision. In principle the energy saved by computing at this minimal level of precision can be redeployed to increase the dynamic range of the model (i.e. its resolution)
into cloud resolved scales. The realisation of this paradigm will require close collaboration with computer hardware developers.