An interpretable framework for tornado loss estimation using parametric surrogate models and dual-objective adaptive smart sampling

Tornadoes pose a significant threat to communities due to their sudden nature, often causing widespread destruction, economic losses, and fatalities. Although the existing tornado damage models provide a rigorous damage assessment methodology, they are limited to a few archetypes and cannot capture the diverse loss patterns across different building geometric configurations. Additionally, these models lack the generalizability needed for use by practitioners and various stakeholders. This study addresses these limitations by developing parametric loss models for one-story residential buildings with gable and hip roofs, demonstrating a methodology that can be generalized to other building configurations. This study employs a previously developed physics-informed probabilistic methodology along with ASCE7-22 tornado guidelines and probabilistic approaches to generate loss curves for buildings with various attributes. To provide vulnerability models in the simplest and most practical form, mathematically interpretable representations of the loss curves were targeted. A representative subset of building configurations was selected using a dual-objective adaptive smart sampling method, ensuring efficient coverage of input and output variable spaces. Various statistical formulations served as surrogates, with linear regressions established between the building properties and the surrogate parameters. While all models showed acceptable accuracy, scaled sigmoid and scaled Weibull CDF formulations were identified as the best representation for vulnerability models of residential buildings with gable and hip roofs, respectively. The result is a simple and concise framework that is applicable to a wide range of stakeholders in the built environment.