Abstract
This thesis develops a novel data-driven framework for rapid vulnerability and resilience assessment of steel special moment-resisting frame (SMRF) buildings under both single-hazard (earthquake) and multi-hazard (sequential earthquake–flood) scenarios. Emphasising methodological innovation, the work extends and integrates established techniques to create new capabilities for structural engineering. A custom meta-database of 30,000 SMRF structures (covering low-, mid-, and high-rise archetypes) is generated to capture uncertainties in geometry, material properties, and nonlinear behaviour. These models obey key design constraints (e.g., strong-column–weak-beam) and were validated against benchmark pushover curves to ensure realism and diversity. To improve seismic input selection for nonlinear time history analysis (NLTHA), a modified Conditional Mean Spectrum (CMS) procedure is proposed, introducing period-specific weighting factors and new cost functions. This enhancement significantly reduces record-to-record variability in structural response (e.g., drift demand), yielding more consistent surrogate-model training data.For the earthquake-only scenario, an optimised machine-learning (ML) surrogate model is developed to predict peak engineering demand parameters (EDPs), such as maximum inter-storey drift ratio (MIDR) and floor acceleration, across the SMRF design space. Among 13 algorithms evaluated, CatBoost was selected for its superior accuracy (validation R² = 0.94, outperforming alternatives by 2–8%) and robustness with large, complex datasets. For the sequential earthquake–flood scenario, a fully integrated three-dimensional (3D) simulation coupling OpenSeesPy (structural analysis) and OpenFOAM (computational fluid dynamics) is created to model complex flood–structure–earthquake interactions, including out-of-plane hydrodynamic forces on building frames. High-fidelity simulations across multiple inundation depths and ground motions provide “ground truth” data to train a Stacked Attention-based LSTM (Stack-AttenLSTM) surrogate model. This deep learning model incorporates both static structural features and dynamic loading time-histories, capturing hazard sequencing effects that traditional surrogates overlook. The trained multi-hazard surrogate achieves high predictive accuracy (R² > 0.87 for MIDR), matching the fidelity of NLTHA while reducing computation time from days to seconds. Finally, the surrogate models are deployed in a user-friendly web application, enabling near-real-time estimation of structural response and resilience metrics. As a demonstration, the tool computes a Resilience Index (Ri) and functionality loss curves for case-study buildings, using FEMA P-58 and REDi framework data to quantify repair time and recovery.
The results demonstrate that the proposed framework achieves predictive accuracy comparable to advanced numerical simulations while accelerating the analysis by up to four orders of magnitude. This represents a transformative step toward performance-based resilient design and rapid post-disaster decision support. The key contributions of this work include novel algorithms for ground-motion selection, an integrated multi-hazard simulation workflow, advanced machine-learning surrogate models that collectively advance the state of the art in structural vulnerability assessment, and the development of a web-based tool for evaluating single- and multi-hazard building vulnerability.
| Date of Award | 29 Apr 2026 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Imrose Bin Muhit (Supervisor), Nashwan Dawood (Supervisor) & Annalisa Occhipinti (Supervisor) |