Mitigating hailstorm impacts: an integrated approach using Machine Learning and physical simulations

Hailstorms and other convective events present significant societal and economic risks, with challenges in assessing these risks due to inconsistent data. The main goal of our research is to help mitigate the impacts of hailstorms by generating high-resolution risk maps. While high-resolution simulations with models like the Weather Research and Forecast (WRF) are key for generating risk maps, these models are computationally expensive. To address this challenge, we have developed a machine learning model, the Convective Days Detector (CDD), which uses ERA5 reanalysis data to identify the days when the probability of deep convection is high. The CDD serves as a starting point for hail event simulations, reducing the number of days to simulate to essential ones, and thereby reducing computational costs. 

We will provide a comprehensive analysis of the CDD, discussing its performance in detecting convective events and the process we followed to refine its capabilities. Our analysis began with a detailed examination of essential instability variables Convective Inhibition (CIN) and Convective Available Potential Energy (CAPE) from the ERA5. This analysis involves comparing different packages for calculating CAPE and CIN with ERA5 data. Despite finding differences, particularly in high CAPE values, we concluded that these discrepancies are not significant for a machine learning model. Therefore, we use ERA5 CAPE due to its importance in detecting convective events and opted not to use CIN in our model due to some inconsistencies we found. Following this, we performed a feature analysis to further refine our model, reducing the number of variables used to eight essential ones for detecting convection. We validated the model to ensure its functionality and introduced a spatial error margin to assess spatial inaccuracies in convective day detection. This validation process involved verifying the performance of the model against known convective events and evaluating its ability to accurately identify convective days. 

Performing high-resolution WRF simulations on days with high deep convection likelihood, we aim to provide a more detailed understanding of the dynamics and impacts of extreme weather events. We will discuss how we plan to use WRF to simulate the physical processes of these events and how these simulations can contribute to the generation of more accurate risk maps. By combining insights from meteorological models, machine learning algorithms, and risk mapping, this research aims to provide a comprehensive framework for understanding and predicting and minimise the socio-economic impact of hailstorms across diverse sectors.