Can less precise models yield more accurate forecasts of weather and climate?
LOCATION
Edgbaston
Birmingham
West Midlands
B15 2TT
United Kingdom
SPEAKER | Dr Tobias Thornes, Atmospheric Oceanic and Planetary Physics, University of Oxford
ABSTRACT | Given that the climate is changing and extreme weather is becoming increasingly prevalent, the need for accurate forecasts of weather and climate is more pressing than ever. One of the key constraints on the quality of forecasts is the resolution and complexity of the numerical models used to inform them, which are themselves constrained by how much computer power is available and affordable to forecast centres. But much energy may be wasted by carrying out all calculations in standard 64-bit 'double-precision'. In this talk, a new method to increase the efficiency of forecasts by removing superfluous precision will be described. Results will be presented that provide evidence in favour of the hypothesis that quantities are less accurately known and therefore do not need to be represented as precisely at smaller spatial scales. If hardware capable of solving equations with less precision at smaller scales were to be deployed operationally, the computational cost savings could be considerable, and these savings could be reinvested to produce forecasts of greater resolution, complexity or ensemble size.
SPEAKER | Dr Tobias Thornes, Atmospheric Oceanic and Planetary Physics, University of Oxford
ABSTRACT | Given that the climate is changing and extreme weather is becoming increasingly prevalent, the need for accurate forecasts of weather and climate is more pressing than ever. One of the key constraints on the quality of forecasts is the resolution and complexity of the numerical models used to inform them, which are themselves constrained by how much computer power is available and affordable to forecast centres. But much energy may be wasted by carrying out all calculations in standard 64-bit 'double-precision'. In this talk, a new method to increase the efficiency of forecasts by removing superfluous precision will be described. Results will be presented that provide evidence in favour of the hypothesis that quantities are less accurately known and therefore do not need to be represented as precisely at smaller spatial scales. If hardware capable of solving equations with less precision at smaller scales were to be deployed operationally, the computational cost savings could be considerable, and these savings could be reinvested to produce forecasts of greater resolution, complexity or ensemble size.