Improving Early-warning Tornado Forecast Based on Environmental Variables Using Symbolic Deep Learning

Tornadoes, considered as one of the severe convective weather systems, cause significant human and economic losses annually. Detecting and predicting the occurrences and magnitude of tornadoes remain challenging due to their rapid evolution on a small spatiotemporal scale. Understanding the relationship between atmospheric/environmental variables and tornado occurrence is critical for developing an early warning and forecast system, as well as for improving our understanding of tornado genesis. Machine learning (ML) and artificial intelligence (AI) have demonstrated their powerful ability to solve nonlinear systems, aided by burgeoning availabilities of observational data and computational resources. However, AI models are conventionally perceived as 'black boxes,' which compromises their full trustworthiness among weather forecasters and decision-makers. To address these concerns, we propose the integration of symbolic regression with Artificial Nural Network (ANN) to explore analytic expressions following physical properties. In doing so, we aim to improve early-warning tornado forecasts and enhance our understanding of tornado genesis using atmospheric/environmental variables. Ultimately, we propose novel tornado-alerting indicators, including the tornado detection and intensity warning, to show their superiority over traditional indicators such as the significant tornado parameter (STP) and the energy helicity index (EHI).