The skill of weather forecasts has improved tremendously over the past three decades. Seven day forecasts today are as skillful as five day forecasts only twenty years ago. This increase in skill is due to systematic improvements in our models of the Earth-System (the atmosphere, ocean, land, biosphere and cryosphere), an expanded network of observing instruments, such as satellites, and more sophisticated techniques for processing the data from those instruments (“data assimilation”).
Underlying all of these improvements has been an exponential increase in the computational power of our supercomputers. However, we cannot rely on this exponential increase to continue indefinitely, as we reach the limits of what is possible with traditional computing hardware. During my DPhil, I have been exploring ways to use novel computing hardware to improve the efficiency of our weather forecasting systems, and therefore the skill of our forecasts.
I have found that our weather models and data assimilation algorithms can be run at a much lower precision than previously thought. Lower precision computations sacrifice mathematical exactness for increased computational speed. Therefore, if we lower precision we can afford to run more accurate model simulations, which would normally require a larger computer. My research could allow operational weather centres, such as the UK Met Office, to boost the skill of their forecasts at a minimal cost.
Sam Hatfield is a DPhil student in Physics in the group of Professor Tim Palmer, who is a Professorial Fellow of Jesus College.