Advancing seismic capacity curve predictions with a meta-modeling framework for structural systems

Predicting accurate seismic capacity curves for reinforced concrete (RC) bridges is essential for assessing seismic performance and resilience, particularly in earthquake-prone regions such as the Pacific Northwest.The Cascadia Subduction Zone megathrust earthquake presents a critical threat to such infrastructure, yet traditional approaches often fail to capture the nonlinearities and high-dimensional relationships between structural parameters and seismic responses.While single-model machine learning (ML) techniques have shown promise, their limited robustness and generalizability hinder their effectiveness.To address these challenges, this research introduces a stacked machine learning framework to enhance the prediction of capacity curves, specifically using column drift ratios as a key measure of structural performance under seismic loading.The proposed framework systematically evaluates various regression models to construct an optimized meta-model by selecting the bestperforming algorithms based on predictive accuracy and generalizability.Incremental Dynamic Analysis (IDA) is performed on randomized RC bridge samples, with input parameters systematically varied using the Latin Hypercube Sampling technique to generate a diverse training dataset.Ground motion suites tailored to Cascadia-specific long-duration seismic characteristics ensure the framework's regional applicability.Advanced ensemble and boostingbased regression algorithms, such as Random Forest and Extreme Gradient Boosting, are tested and the top-performing models are combined into the meta-model.This stacked approach addresses the limitations of single-model techniques by improving robustness, accuracy, and the ability to generalize across diverse structural configurations.The ML models predict capacity curves based on column drift ratios, providing precise and actionable insights into structural performance.To ensure interpretability, Explainable AI (XAI) techniques, such as SHapley Additive exPlanations (SHAP), are used to rank the importance of input parameters, offering actionable insights for seismic design and retrofit strategies.This stacked machine learning approach offers several advantages, including the ability to model complex, nonlinear relationships and produce accurate capacity curve predictions for RC bridges.Results demonstrate that the framework enhances the reliability of seismic risk assessments and reduces computational demands, empowering engineers and decision-makers to design, evaluate, and retrofit infrastructure with greater confidence.By overcoming the limitations of existing methodologies, this research highlights the transformative potential of stacked machine learning in advancing seismic risk assessment practices and strengthening infrastructure resilience.