Revisiting Machine Learning Approaches for Short- and Longwave Radiation
Inference in Weather and Climate Models, Part II: Online Performance
Abstract
This paper continues the exploration of \gls{ml}
parameterization for radiative transfer for the
\gls{icon}. Three \gls{ml} models,
developed in Part I of this study, are coupled to
\gls{icon}. More specifically, a UNet model and a
bidirectional \gls{rnn} with
\gls{lstm} are compared against a random forest. The
\gls{ml} parameterizations are coupled to the
\gls{icon} code that includes OpenACC compiler
directives to enable \glspl{gpu} support. The coupling
is done through Infero, developed by ECMWF, and PyTorch-Fortran. The
most accurate model is the bidirectional \gls{rnn} with
physics-informed normalization strategy and heating rate penalty, but
the fluxes above 15\,km height are computed with a
simplified formula for numerical stability reasons. The presented setup
enables stable aquaplanet simulations with \gls{icon}
for several weeks at a resolution of about 80\,km and
compare well with the physics-based radiative transfer solver ecRad.
However, the achieved speed up when using the emulators and the minimum
required memory usage relative to the \gls{gpu}-enabled
ecRad depend strongly on the \gls{nn} architecture.
Future studies may explore physics-constraint emulators that predict
heating rates inside the atmospheric model and fluxes at the top.