Performance evaluation of machine learning algorithms for predicting liquefaction-induced lateral displacement.

Area of Science:

• Geotechnical Engineering
• Earthquake Engineering
• Machine Learning Applications in Civil Engineering

Background:

• Accurate prediction of liquefaction-induced lateral displacement is crucial for seismic risk assessment and resilient infrastructure design.
• Conventional analytical methods struggle to reliably address the complexities of predicting these displacements.
• Existing models often lack the precision required for effective mitigation strategies.

Purpose of the Study:

• To model and investigate liquefaction-induced lateral displacements using advanced machine learning algorithms.
• To compare the predictive performance of XGBoost, CatBoost, and AdaBoost against established methods.
• To identify the most influential parameters affecting lateral displacement predictions.

Main Methods:

• Utilized 247 in-situ free-face ground condition case studies of post-liquefaction.
• Employed XGBoost, CatBoost, and AdaBoost algorithms for predictive modeling.
• Evaluated model performance using metrics such as R², r, MAE, MSE, RMSE, RSR, and NSE, and compared with literature models.

Main Results:

• The XGBoost model exhibited the best prediction performance, achieving R² = 0.9905 (training) and R² = 0.9251 (testing).
• XGBoost outperformed other developed models and literature models like Gaussian process regression and artificial neural networks.
• Sensitivity analysis indicated that T₁₅ was the most sensitive parameter influencing lateral displacement.

Conclusions:

• XGBoost offers a superior approach for predicting liquefaction-induced lateral displacements compared to conventional and other machine learning methods.
• The developed XGBoost model enhances the accuracy of seismic risk assessment and infrastructure design.
• Understanding parameter sensitivity, such as T₁₅, aids in refining predictive models and mitigation strategies.