PhD defense by Peter Ukkonen

Thesis title: "Substantially faster radiation schemes for weather and climate models by using machine learning and code optimization"

Thesis abstract:

Radiative transfer parameterizations are physically fundamental components of weather and climate models, and often represent a computational bottleneck in climate models. At the same time, the amount of energy needed to run these models is growing alongside a push towards ever-increasing resolution and physical realism. It is therefore clear that the trade-off between computational efficiency and accuracy in radiation parameterizations is very important and needs to improve. This PhD study aims to contribute to this worthy goal by approaching it from two angles: using approximative but faster machine learning methods to replace physical radiation schemes or their components, and two, improving the efficiency of existing schemes by using code restructuring techniques.

Here it is shown that by replacing only one component of a radiation parameterization with a neural network (NN), significant improvements in runtimes can be achieved without any considerable loss of accuracy. This component, known as the gas optics scheme, computes the optical properties of the gaseous atmosphere. Combining the NN gas optics with a refactored radiative transfer solver, a modern radiation scheme (RTE+RRTMGP) was made 2-3 times faster. By implementing the NN gas optics in the "ecRAD" radiation scheme used in a leading state-of-the-art weather model, the Integrated Forecast System (IFS), it was demonstrated that compared to using the original gas optics, the NN emulator does not significantly impact the model climate and speeds up ecRAD by roughly a third.

This PhD thesis also contributes to more accurate emulation of atmospheric radiative transfer (the full radiation scheme) by development of a novel method based on recurrent neural networks (RNNs), the structure of which more closely reflects the physics of radiative transfer. Shortwave fluxes and heating rates can be predicted with far greater accuracy compared to using standard feed-forward NNs, while also requiring several orders of magnitude fewer model parameters.

Finally, a significant code restructuring of ecRAD was carried out. The focus was on improving the efficiency of a radiative transfer solver capable of representing the 3-D radiative effects of clouds (SPARTACUS). These 3-D effects are currently ignored in all weather and climate models. The computational cost of optimized SPARTACUS, when combined with an advanced new gas optics scheme with a smaller spectral resolution, is actually less than the operational radiation code in the IFS! The impact of these results should be significant, assuming some remaining issues with numerical instability when running SPARTACUS in single precision can be resolved.

Chair of commettee: Jens Hesselbjerg Christensen

Censors: Quentin Libois and Peter Dueben

Supervisor: Eigil Kaas