Natural Hazards Research Summit 2024: Theory-Guided Statistical Inference Framework for Advancing Knowledge Discovery from Post-Windstorm Engineering Assessments

This study summarizes advances towards a theory-guided statistical inference framework that blends physics-based modeling with data-driven analysis methods to infer causal relationships in windstorm performance of low-rise buildings.
Data from 10 different windstorms were curated into a single, unified testbed of building performance assessments of single-family residential structures (predominately wood-frame) creating the WindStorm Performance Dataset (WiSPD). Each record contains an estimated peak 3-sec gust wind speed, common building attributes (e.g., roof shape, number of stories, year built), surrounding terrain conditions, overall wind and surge damage ratings, and component-level damage percentages (CLDP). In parallel, a framework has been developed for separately quantifying the resistance-based and aerodynamic vulnerability of wood-frame housing. System-level structural vulnerability is evaluated by transforming the wind uplift load path into a chain of series elements and estimating relative failure probabilities by Monte Carlo simulation, thereby reducing a wide array of structural factors to a few critical parameters that most influence wind resistance. Aerodynamic factors, such as roof type, roof slope, and eave height, are similarly reduced to a vulnerability index by means of a pressure prediction model trained on wind tunnel data, which estimates mean wind uplift intensity across all wind directions.
Bayesian analysis techniques provide methods which can integrate prior knowledge, sequential learning, and uncertainty quantification into a single framework, thus creating the opportunity to combine the empirical data from the WiSPD and knowledge gained from theory-guided physics-based modeling. Starting with a Bayesian ordinal logistic regression (B-OLR), this technique explores the relationships between parameters contained in the WiSPD and quantifies the uncertainty of each parameter’s influence. However, using the empirical data alone, the accuracy of this model remains low, highlighting the need for additional knowledge. In future iterations of this analysis the WiSPD testbed will be expanded to integrate the physics-informed vulnerability factors and will incorporate prior knowledge in the form of a physics-based synthetic dataset developed from the NHERI SimCenter’s R2D tool. The frameworks presented herein provide a promising tool for incorporating multiple sources of data, blending physics-based models and data-driven techniques, capturing established theory and knowledge. Ultimately, this new knowledge will then be used to inform a Bayesian Network for continuous updating and causal inference for future windstorms and educating communities. By combining theory, the principles established in the Davenport Wind Loading Chain, and the results of the previously mentioned analysis techniques, the Bayesian Network provides a framework to formalize a holistic understanding of the factors driving structural performance in windstorms for use in education, causal inference, and policy-making endeavors.