Stack-AttenLSTM: A surrogate deep learning model for sequential earthquake-flood structural response assessment of steel buildings

Accurately assessing the vulnerability of critical building portfolios is fundamental for regional risk management and decision support, especially in regions facing sequential earthquake–flood events exacerbated by climate change. Such compound disasters pose severe environmental challenges, yet current practice lacks reliable surrogate models to rapidly predict structural response and damage-relevant demand parameters under combined seismic and flood loading. This study addresses that gap by introducing a soft computing approach, the Stacked Attention-based Long Short-Term Memory network (Stack-AttenLSTM), to efficiently predict key structural response quantities under sequential earthquake–flood hazards. In this framework, structural vulnerability is interpreted in a performance-based sense, whereby hazard-induced response metrics serve as proxies for damage susceptibility rather than direct loss estimation. The surrogate model predicts key engineering demand parameters (EDPs), including the maximum inter-storey drift ratio (MIDR), maximum floor acceleration (MFA), and maximum base shear (MBS), which are widely used indicators of structural damage and vulnerability. To develop this model, a large-scale meta-database comprising 30,000 steel special moment-resisting frame (SMRF) buildings is first generated to capture structural variability across low-, mid-, and high-rise typologies, from which a representative subset is selected for detailed high-fidelity three-dimensional (3D) nonlinear time-history analyses (NLTHA) under sequential earthquake–flood loading and surrogate model training. Flood loading is represented using computational fluid dynamics (CFD) simulations with a dam-break–type inflow condition, employed as a conservative hydrodynamic proxy to study flow-induced forces under extreme inundation scenarios. Multiple Stack-AttenLSTM architectures are trained and evaluated, and the final model is selected for its optimal balance of predictive accuracy and computational efficiency, enabling rapid yet reliable response prediction. The proposed model achieves high predictive accuracy, with coefficients of determination (R²) approaching 0.88 and low error metrics across all hazard scenarios, demonstrating its effectiveness for rapid multi-hazard vulnerability assessment. Although explicit fragility or loss models are not derived, the Stack-AttenLSTM framework is suitable for integration with early warning systems and digital twin platforms, enabling real-time monitoring, improved uncertainty management, and proactive disaster response.