Scholarlink

<p>Lateral spreading induced by earthquake shaking is a major cause of damage to infrastructure and lifelines. With rapidly increasing urbanisation in seismic regions, accurate prediction of lateral spreading has become critical for resilient infrastructure design and seismic hazard mitigation. The present study develops and evaluates five supervised machine-learning models, i.e., Gaussian Process Regression (GPR), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), Categorical Boosting (Cat Boost), and Extreme Gradient Boosting (XG Boost), to predict the lateral displacement. Six geotechnical and seismic input parameters are used in the model development. The input parameters include earthquake magnitude, horizontal source distance from the site, horizontal ground slope, average fine content, liquefiable soil layer thickness, and average mean grain size. This study also examines the sensitive regions which are more sensitive to liquefaction and lateral spreading. Developed models are then tuned systematically using Bayesian, Grid search, and randomised search, combined with 5-fold cross-validation. The result of the study shows that the model performance improves significantly after tuning the hyperparameters, and the relative ranking of the models changes compared to the baseline result. The optimised Cat Boost model significantly improves their prediction accuracy and becomes the best predictive model with an RMSE of 0.006 and an R<sup>2</sup> of 0.994 among all the models. The performance of Cat Boost exceeds the GPR model, which initially performed best before optimisation. The reliability and interpretability of the model are also evaluated using diagnostic plots, error distribution analysis, and feature importance evaluation. The results confirm that careful hyperparameter optimisation is essential for producing stable and physically meaningful predictions. The study also identifies that the highest lateral displacements occur in the highly seismic regions of Darbhanga, Purnia, and Muzaffarpur districts. The proposed modelling framework can support improved prediction and assessment of liquefaction-induced ground deformation.</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

Optimised machine learning predictions of liquefaction-induced lateral spreading in alluvial deposits

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Dilip Kumar,
  • Sunita Kumari

摘要

Lateral spreading induced by earthquake shaking is a major cause of damage to infrastructure and lifelines. With rapidly increasing urbanisation in seismic regions, accurate prediction of lateral spreading has become critical for resilient infrastructure design and seismic hazard mitigation. The present study develops and evaluates five supervised machine-learning models, i.e., Gaussian Process Regression (GPR), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), Categorical Boosting (Cat Boost), and Extreme Gradient Boosting (XG Boost), to predict the lateral displacement. Six geotechnical and seismic input parameters are used in the model development. The input parameters include earthquake magnitude, horizontal source distance from the site, horizontal ground slope, average fine content, liquefiable soil layer thickness, and average mean grain size. This study also examines the sensitive regions which are more sensitive to liquefaction and lateral spreading. Developed models are then tuned systematically using Bayesian, Grid search, and randomised search, combined with 5-fold cross-validation. The result of the study shows that the model performance improves significantly after tuning the hyperparameters, and the relative ranking of the models changes compared to the baseline result. The optimised Cat Boost model significantly improves their prediction accuracy and becomes the best predictive model with an RMSE of 0.006 and an R2 of 0.994 among all the models. The performance of Cat Boost exceeds the GPR model, which initially performed best before optimisation. The reliability and interpretability of the model are also evaluated using diagnostic plots, error distribution analysis, and feature importance evaluation. The results confirm that careful hyperparameter optimisation is essential for producing stable and physically meaningful predictions. The study also identifies that the highest lateral displacements occur in the highly seismic regions of Darbhanga, Purnia, and Muzaffarpur districts. The proposed modelling framework can support improved prediction and assessment of liquefaction-induced ground deformation.