Quantifying the relationship between extratropical cyclones' intensity measures and their background state: systematic exploration of a baroclinic wave simulation ensemble

Extreme extratropcial cyclones (ETC) are associated with heavy precipitation, and strong winds causing damage to infrastructure, or diverse economic losses. They can be characterised by a set of variables or diagnosis named intensity measures. Based on them, meteorologists are able to study the intricate relationship between dynamical features and impacts of ETCs. However, considerable additional research is required to improve our understanding of the relationship between ETCs' intensity and the the background state they develop in. Our baroclinic wave simulation setup implemented in OpenIFS 43r3 has shown the possibility to create stable and flexible background states able to run with moisture and full physics. Moreover, 7 parameters can be easily varied to produce a vast array of different background states. By varying these parameters, an ensemble of 6,500 baroclinic waves are simulated using OpenIFS@home, a version of OpenIFS that runs within a volunteer computing framework. In these cases, the developing ETCs are physically realistic with poleward motion, upstream and downstream development and sensible minimum mean sea level pressure.

This study proposes a Machine Learning based and systematic approach to link the 6,500 background states with their developing ETCs. Each ETC is isolated and tracked. A total of 75 features are extracted from tracked ETCs for each case. A Random Forest Regressor (RFR) is use to predict each 13 intensity measures with 5 background features. One of the properties of the RFR is its ability to rank its input features during the training. As a result, this embedded feature selection allow to quantify the strength of relationship - called feature importance. For example, the feature importance between the initial average temperature with the resulting accumulated precipitation, or the horizontal temperature gradient at 300hPa with the maximum relative vorticity at 850hPa can be estimated. The proposed methodology is able to (1) predict 13 intensity measure, (2) link them to 5 background features, and (3) reduce the training dataset by filtering the ETCs to the most intense. To stabilise the feature selection, a bootstrapping-based approach has been implemented. Using the distributed nature of the workflow, the whole ensemble of 6,500 baroclinic waves is processed within 1.5 days on 40 cores and the computational time reduces linearly with the number of cores.

With the exception of the storm severity index, the RFR is able to predict the intensity measures with a coefficient of determination between 0.65 and 0.92. Moreover, this study demonstrates an increase feature importance of the upper-troposphere baroclinicity as the training dataset is reduced to the most intense ETCs. Concurrently, the importance of the lower-tropospheric baroclinicity decreases. The feature importance of friction, initial relative humidity, initial laps rate and average surface virtual temperature stays constant. Future work will include the use of Deep Learning Regressor and wrapped feature selection in order to validate and extend the main result of this study.