Samuel Hatfield, University of Oxford
The use of inexact hardware in data assimilation for improved weather and climate prediction
Coauthors
Tim Palmer, Peter Düben, Aneesh SubramanianAbstract:
Recent studies have shown that numerical precision can be reduced significantly within atmospheric models with no significant increase in forecast error. Such a reduction in precision frees up computational resources that can be reinvested in other parts of the model, such as the grid resolution or ensemble size, thereby improving the quality of weather and climate predictions. We now extend this research effort to data assimilation and present preliminary results of a study of trade-offs between numerical precision and performance in a toy model for atmospheric dynamics (the Lorenz '96 model) with an ensemble Kalman filter. We consider several ways that computational savings can be reinvested, notably the ensemble size and the number of observations assimilated.
To include scale interactions analogous to those observed in the atmosphere, we consider a three-level version of the Lorenz '96 model that was recently developed in our working group. In this setup, the three-level model can serve as a surrogate for truth, while observations of this truth are assimilated into a two-level model with parametrised small scale dynamics. The model is run on an emulator of an inexact processor, which supports arbitrary precision variables. In this way, results may help in the design of the next generation of high performance chips within the field of data assimilation, by identifying the components of the data assimilation cycle tolerant to reduced precision.