Scholarlink

<p>Seismic hazard assessment has traditionally relied on a limited set of parameters such as earthquake magnitude, epicentral distance, and site conditions. However, these measures often fail to capture the complex interactions governing seismic ground motion. This study applies a data-driven approach to evaluate the predictive power of integrated intensity measures, specifically Arias Intensity (IA) and Cumulative Absolute Velocity (CAV), for estimating peak ground-motion parameters (PGA and PGV). A dataset of 4101 three-component strong-motion records from the Southern Turkey seismic sequence was analyzed using a diverse set of machine learning models. Results show that IA and CAV consistently outperform traditional source and site parameters. For the combined (component-pooled) dataset, where the E, N, and Z components were analyzed together, very high predictive performance was achieved, with Gradient Boosting providing the best accuracy for PGA (R² = 0.975) and K-Nearest Neighbors yielding the best performance for PGV (R² = 0.883), closely followed by Random Forest (R² = 0.882). SHapley Additive exPlanations (SHAP) analysis confirmed the dominant influence of energy-based measures on predictions, while conventional descriptors (e.g., magnitude, distance, depth, V<sub>S</sub>30, and spatial metadata) provided comparatively smaller marginal contributions once waveform-derived measures were included. These findings highlight the value of energy-based measures as robust proxies for complex seismic effects, offering improved predictive accuracy and actionable insights for seismic hazard assessment.</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

Explainable machine learning for ground motion modeling: predictive dominance of energy-based measures in the Southern Turkey seismic sequence

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Mohammad Reza Bagerzadeh Karimi

摘要

Seismic hazard assessment has traditionally relied on a limited set of parameters such as earthquake magnitude, epicentral distance, and site conditions. However, these measures often fail to capture the complex interactions governing seismic ground motion. This study applies a data-driven approach to evaluate the predictive power of integrated intensity measures, specifically Arias Intensity (IA) and Cumulative Absolute Velocity (CAV), for estimating peak ground-motion parameters (PGA and PGV). A dataset of 4101 three-component strong-motion records from the Southern Turkey seismic sequence was analyzed using a diverse set of machine learning models. Results show that IA and CAV consistently outperform traditional source and site parameters. For the combined (component-pooled) dataset, where the E, N, and Z components were analyzed together, very high predictive performance was achieved, with Gradient Boosting providing the best accuracy for PGA (R² = 0.975) and K-Nearest Neighbors yielding the best performance for PGV (R² = 0.883), closely followed by Random Forest (R² = 0.882). SHapley Additive exPlanations (SHAP) analysis confirmed the dominant influence of energy-based measures on predictions, while conventional descriptors (e.g., magnitude, distance, depth, VS30, and spatial metadata) provided comparatively smaller marginal contributions once waveform-derived measures were included. These findings highlight the value of energy-based measures as robust proxies for complex seismic effects, offering improved predictive accuracy and actionable insights for seismic hazard assessment.