Identifying controls on extra-tropical cyclone intensity using baroclinic wave simulations and machine learning

Extra-tropical cyclones (ETCs) are a key part of the mid-latitude circulation and can lead to hazardous weather such as heavy precipitation and strong winds. However, how the intensity of extra-tropical cyclones will change in the future remains somewhat uncertain. This is partially due to the “tug-of-war” between the predicted increase in the upper-level temperature gradient, which would theoretically increase ETC intensity, and the predicted decrease in the lower-level temperature gradient which would decrease the intensity of ETCs. Additionally how diabatic processes will influence the intensity of ETCs in the future also remains uncertain. To quantify the relative contribution of these different factors, we adopt an idealised modelling approach. We have performed a large (6500 members) ensemble of baroclinic life cycle simulations with OpenIFS, which is a version of ECMWF’s Integrated Forecast System. Each ensemble member has a different initial state which is generated by varying 7 parameters: surface temperature; surface relative humidity; jet width, height and strength; lapse rate; and surface roughness. We objectively track the ETCs which develop in each simulation and for each ETC we compute 13 intensity metrics, such as maximum vorticity, wind footprint, a storm severity index and accumulated precipitation. A Random Forest Regressor (RFR) is then use to separately predict each of the 13 intensity measures using 5 features describing the initial state as input. One of the properties of RFR is its ability to rank its input features during the training. As a result, this embedded feature selection allow us to quantify the strength of relationship - called feature importance – and thus identify the aspects of the initial state which have the strongest influence on the resulting ETCs intensity. 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. Wind-based measures are better predicted than all of the precipitation intensity measures. When all ensemble members are considered, we find that the low-level temperature gradient is the most important feature explaining ETC intensity, closely followed by the upper-level temperature gradient. However, when only the most extreme ETCs are considered, the upper-level temperature gradient becomes the dominant feature. These results, along with a discussion of their implications, will be presented.