Scholarlink

<p>Accurate prediction of seismic demand is central to performance-based earthquake engineering. However, it remains unknown whether a single, interpretable, and uncertainty-aware hybrid model can capture the intensity-dependent transitions that shape structural response under contrasting ground-motion conditions. This work establishes a hybrid machine-learning framework that integrates optimized base learners through blending for base shear and stacking for interstory drift ratio to deliver high-resolution, physically coherent response predictions. The model was trained on 500 elastic time-history analyses generated by applying 20 recorded earthquakes to 25 reinforced-concrete frames and evaluated across three excitation contexts: a general dataset, near–far fault classifications, and low–high spectral acceleration partitions. The framework achieved strong predictive performance (R² up to 0.98 for base shear and 0.97 for drift). At the same time, interpretability analyses identified consistent governing predictors, including first-mode period, spectral acceleration, and strong-motion intensity measures. They revealed apparent shifts in the influence of predictors as shaking intensity increased. Uncertainty quantification showed that bootstrap intervals had empirical coverage exceeding 90%, whereas parametric intervals did not adequately cover high-intensity and near-fault motions. This indicates that the error structure is asymmetric and heteroscedastic. Taken together, these findings establish a fast, interpretable, and physically grounded surrogate for dynamic analysis, enabling intensity-sensitive seismic-demand estimation, regional scenario evaluation, and accelerated post-earthquake decision support within performance-based engineering.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid machine learning surrogates for seismic demand prediction with SHAP-based interpretability and uncertainty quantification

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Ali Kaveh,
  • Alireza Amiri

摘要

Accurate prediction of seismic demand is central to performance-based earthquake engineering. However, it remains unknown whether a single, interpretable, and uncertainty-aware hybrid model can capture the intensity-dependent transitions that shape structural response under contrasting ground-motion conditions. This work establishes a hybrid machine-learning framework that integrates optimized base learners through blending for base shear and stacking for interstory drift ratio to deliver high-resolution, physically coherent response predictions. The model was trained on 500 elastic time-history analyses generated by applying 20 recorded earthquakes to 25 reinforced-concrete frames and evaluated across three excitation contexts: a general dataset, near–far fault classifications, and low–high spectral acceleration partitions. The framework achieved strong predictive performance (R² up to 0.98 for base shear and 0.97 for drift). At the same time, interpretability analyses identified consistent governing predictors, including first-mode period, spectral acceleration, and strong-motion intensity measures. They revealed apparent shifts in the influence of predictors as shaking intensity increased. Uncertainty quantification showed that bootstrap intervals had empirical coverage exceeding 90%, whereas parametric intervals did not adequately cover high-intensity and near-fault motions. This indicates that the error structure is asymmetric and heteroscedastic. Taken together, these findings establish a fast, interpretable, and physically grounded surrogate for dynamic analysis, enabling intensity-sensitive seismic-demand estimation, regional scenario evaluation, and accelerated post-earthquake decision support within performance-based engineering.