Optimal Dimensioning of Retaining Walls Using Explainable Ensemble Learning Algorithms.

Area of Science:

• Civil Engineering
• Computational Engineering
• Machine Learning

Background:

• Reinforced concrete retaining walls are crucial structures.
• Optimizing their dimensions is key to minimizing construction costs.
• Traditional design methods can be time-consuming and may not explore all optimal solutions.

Purpose of the Study:

• To develop and compare machine learning models for predicting optimal dimensions of reinforced concrete retaining walls.
• To minimize construction costs through accurate dimension prediction.
• To enhance the efficiency and applicability of machine learning in structural engineering design.

Main Methods:

• Utilized Random Forest, XGBoost, CatBoost, and LightGBM algorithms for predictive modeling.
• Generated a comprehensive dataset using the Harmony Search (HS) algorithm.
• Employed SHapley Additive exPlanations (SHAP) to visualize feature influence on optimal dimensions.
• Evaluated model accuracy using R-squared, RMSE, and MAE metrics.

Main Results:

• Achieved a high prediction accuracy with an R2 score of 0.99 on the test set.
• LightGBM demonstrated the fastest computational speed (6.17s).
• CatBoost yielded the highest prediction accuracy among the tested algorithms.
• SHAP analysis provided insights into the interdependence of design features and optimal dimensions.

Conclusions:

• Machine learning algorithms can effectively supplement traditional design procedures for reinforced concrete retaining walls.
• The proposed methodology enables the creation of larger datasets, improving ML model accuracy and applicability.
• Accurate dimension prediction leads to significant construction cost savings and design optimization.