Data assimilation, the process by which atmospheric data is combined with atmospheric models, is essential for skillful weather forecasts. Inserting data into models allows us to characterize more accurately the weather state at the start of the forecast, thereby extending the time in which the forecast is useful. However, data assimilation is a very computationally expensive process, and often costs as much as the actual weather forecasts. One way to reduce the cost is to lower the precision of the data assimilation computations. Lower precision computations use fewer bits to produce the answer to a calculation, and are therefore computationally cheaper. Lowering precision also introduces errors, but these errors may be acceptable, given that our models and observations are imperfect.
I will present some results from my PhD research on the subject of precision in data assimilation. I will show how lowering the precision of the data assimilation algorithm affects the quality of the output. I will demonstrate that the lowest precision that you can use is related to the overall quality of the model and the observations – the better the model, the more important precision becomes. Additionally, I will show that, by reusing the computational resources that we save when lowering precision, we can actually improve the data assimilation product and – ultimately – the skill of weather forecasts.