Data-driven ensemble learning approach for optimal design of cantilever soldier pile retaining walls

Abstract

Cantilever soldier pile retaining walls are used to ensure the stability of excavations. This paper deploys ensemble machine learning algorithms towards achieving optimum design of these structures. A large dataset was developed consisting of 40,569 combinations of pile geometry, external loading, soil properties, and concrete unit cost, with two different values of soil reaction coefficient. Optimum pile diameter that minimizes the total cost of the retaining wall was computed considering the structural load-carrying capacity as the optimization constraint. The dataset was split into training and testing sets at 70% to 30% ratio. The predictive accuracy of the ensemble machine learning models was appraised on the testing dataset using various statistical metrics. Model performance was also evaluated for its ability in predicting the optimum pile diameter. The developed models demonstrated excellent predictive accuracy. Furthermore, the effect of different input variables on the model predictions was explained using the SHapely Additive exPlanations (SHAP) approach. Through the SHAP algorithm, the pile length was identified as the design variable having the most significant effect on the optimum pile diameter. The study demonstrates ensemble learning techniques as a viable alternative to the traditional techniques in the optimum design of cantilever soldier pile retaining walls.