4. Climate - Machine Learning in Climate Research
Understanding climate is one of the most pressing and most complex issues that humanity faces. Modelling of geophysical fluid phenomena is, however, challenging due to dynamic interactions across a wide range of spatial and temporal scales, from molecular to planetary and novel ML based analysis techniques are needed. SMASH’s excellence was recently recognized by the prestigious Puh award and ’Excellent in science 2021’ prize for computer science, specifically for the deep learning coastal flooding model HIDRA, that is now part of the Slovenian Environment Agency forecasting operational pipeline.
4.1 Extreme weather
Climate-related extreme weather events have caused economic losses totalling an estimated €446 billion in the EEA member countries between 1980 and 2019. In Slovenia a single such event can cause millions of euros in damages. ARSO host groups work on extensions and transformations of existing ML algorithms to facilitate better predictions, constrained online with a much larger multi-platform multi-sensor dataset of existing observations, which may be non-regular, noisy and sparse in space and time.
4.2 Role of anthropogenic and natural aerosols and their complex mixtures on the climate
SMASH uses “big-data” methods in the validation of regional climate models to calculate the atmospheric heating of light aerosols. This is needed to investigate the role of fine combustion aerosols and coarse natural mineral dust in their heating/cooling effects, relevant for the validation of climate models and enhancing the predictability in the long-term modelling of different emission reductions.
4.3 Determination of sources of air pollution
Source apportionment is an inverse problem requiring sophisticated approaches to process huge amounts of data. An ideal technique uses ambient measurements with no a-priori information to obtain the sources and their contribution to ambient air pollution. Research is leading to the development of new source apportionment techniques, their validation and application in local and regional air quality abatement strategies.
4.4 Machine learning based long-range atmospheric forecasting
The meteorological community is increasingly using modern ML techniques to improve specific aspects of weather prediction. It is conceivable that some day the data-driven approach will beat the Numerical Weather Prediction (NWP) using the laws of physics, but fundamental breakthroughs are needed. One area where the ML-based approaches show special promise is the Long-Range Atmospheric Forecasting (LRAF). UL FMF group works on devising novel ML- based models for LRAF, identifying sources of long- range atmospheric predictability and evaluating the ML-based forecast skill with respect to the NWP prediction systems.