Joint Seminar: Inexact computers for more accurate forecasts: Are we over-engineering numerical precision in weather and climate models?

In numerical atmosphere and ocean models, values of relevant physical
parameters are often uncertain by more than 100% and weather forecast
skill is decreasing significantly after a couple of days. Still,
numerical operations are typically calculated with 15 significant
decimal digits. If we reduce numerical precision, we can reduce power
consumption and increase computational performance significantly. This
would allow an increase in resolution in weather and climate models and
might be a short-cut to global cloud-resolving modelling.


In co-operations with groups in computing science we study different
approaches to the use of inexact hardware in the hierarchy of models
(from Lorenz ‘95 all the way up to IFS) and estimate possible savings.
Results show that numerical precision can be reduced significantly
within simulations with a three-dimensional dynamical core of an
atmosphere model with no significant increase in model error for both
weather and climate type simulations. If computational cost is reduced
due to the use of inexact hardware, the possible increase in resolution
allows a much stronger reduction of model error compared to the increase
of errors due to reduced precision.


However, there is no doubt that rounding errors will influence numerical
simulations. We study the impact of hardware errors on model dynamics at
different spatial scales as well as interactions between the rounding
error forcing and stochastic forcings of stochastic parametrisation
schemes. We find that a scale selective approach that introduces
significant rounding errors into the computation of small scale dynamics
has huge potential due to the high inherent uncertainty which is present
in these scales due to viscosity, sub-grid-scale variability and
parametrisation schemes. We also find that rounding errors can be hidden
inside the distribution of stochastic forcings of stochastic
parametrisation schemes and that rounding errors can be used to generate
forecast ensembles of similar spread and quality compared to ensembles
based on stochastic parametrisation schemes, at least in idealised test






13:30 h


Bundesstr. 53, room 022/023
Seminar Room 022/023, Ground Floor, Bundesstrasse 53, 20146 Hamburg, Hamburg


Peter Düben, ECMWF


Lorenzo Tomassini

Back to listing