Significant duration prediction of seismic ground motions using machine learning algorithms.

Area of Science:

• Geophysics and seismology
• Computational seismology
• Machine learning applications in earth sciences

Background:

• Predicting seismic motion duration is crucial for earthquake engineering and hazard assessment.
• Existing models often lack accuracy in capturing the complex relationships between seismic parameters.
• The need for advanced computational methods to improve seismic duration prediction is evident.

Purpose of the Study:

• To predict seismic motion significant duration (D5-75, D5-95) using machine learning algorithms.
• To introduce and optimize characteristic parameters for enhanced prediction accuracy.
• To develop and validate novel fusion models for superior seismic duration forecasting.

Main Methods:

• Utilized XGBoost for characteristic parameter optimization, identifying an optimal four-parameter combination.
• Compared prediction performance of Random Forest, XGBoost, BP Neural Network, and SVM algorithms.
• Developed and evaluated two fusion models: stacking and weighted averaging.

Main Results:

• Fusion models (stacking and weighted averaging) significantly improved prediction accuracy and generalization ability compared to single algorithms.
• Optimized parameter combinations enhanced the predictive power of machine learning models.
• Residual analysis confirmed the improved performance and reliability of the fusion models.

Conclusions:

• Machine learning, particularly fusion models, offers a robust approach for predicting seismic motion significant duration.
• The integration of additional parameters like fault top depth and epicenter mechanism parameters refines predictive capabilities.
• The developed fusion models provide a validated, accurate, and rational method for seismic duration prediction, outperforming existing approaches.