Predicting continuous load-settlement trajectories of bored piles: A GA-optimized weighted ensemble framework with explainable AI
- 1University of Transport Technology, Hanoi, Vietnam.
- 2Le Quy Don Technical University, Hanoi, Vietnam.
Abstract
Accurate prediction of bored pile settlement is a core element in reliable foundation design. While machine learning (ML) models demonstrate superior predictive capabilities compared to traditional analytical methods, their black-box nature and tendency to generate non-physical trajectories significantly hinder their practical application in geotechnical engineering. This study proposes a new computational framework that ensures physical consistency by combining Weighted Ensemble learning with SHAP (SHapley Additive exPlanations) interpretation capabilities. The integrated structure incorporates linear regression (LR), Random Forest (RF), XGBoost, and multilayer neural networks (MLP) to capture the complex interaction between piles and soil through a database of 563 field static load test records. A two-stage hybrid optimization strategy, using Random Search combined with a Genetic Algorithm (GA), was implemented to optimize the global weight set. The results showed that the GA-Ensemble model achieved excellent accuracy on the independent test set (R2 = 0.958, NRMSE = 5.27%, RMSE = 0.380, MAE = 0.2322), successfully neutralizing the ladder errors of the tree model to recreate a smooth, continuous, and mechanically compliant load-settlement (P-S) trajectory. Notably, by converting the SHAP value to a real physical scale (millimeters), the study clarified the algorithm’s inference logic. SHAP dependency analysis demonstrated that the model successfully captures the coupling characteristics of elastic compression and geometric nonlinearity by identifying the interaction between external force (P) and free length (Lf). Local diagnostic decompositions precisely extract failure mechanisms, successfully quantifying the “systemic penalty” caused by elastic shortening of structures and foundation puncture even in very stiff soil layers. In short, the proposed method has gone beyond the limits of a purely interpolation tool, establishing a transparent and geomechanically reliable expert system for predicting settlement.
Related
- Predicting Synergistic and Antagonistic Micellization Effects in Surfactant Blends for Optimized Formulation Design.
- Predicting nanoparticle tumor delivery efficiency using explainable machine learning for cancer nanomedicine design
- Hydrothermal-Assisted Extraction of Cinnamomum camphora Essential Oil Optimized by Machine Learning
- Optimized Biomimetic Syntheses of Discorhabdin B and Aleutianamine Drives a Deeper Exploration of Their Anticancer Potential
- Optimized Synthesis and Properties of Dumbbell-Shaped NaYF4:Yb,Tm@NaGdF4:Eu@NH2-MIL-53(Fe) Core–Shell Nanocomposite
BibTeX
@article{giap2026predicting,
title = {{Predicting continuous load-settlement trajectories of bored piles: A GA-optimized weighted ensemble framework with explainable AI}},
author = {Loi Van Giap and Tuong Lai Nguyen and Tuan Anh Pham},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0360945},
url = {https://doi.org/10.1371/journal.pone.0360945},
note = {CC BY},
}
RIS
TY - JOUR
TI - Predicting continuous load-settlement trajectories of bored piles: A GA-optimized weighted ensemble framework with explainable AI
AU - Loi Van Giap
AU - Tuong Lai Nguyen
AU - Tuan Anh Pham
PY - 2026
DA - 2026/10/21
JO - PLOS One
DO - 10.1371/journal.pone.0360945
UR - https://doi.org/10.1371/journal.pone.0360945
AB - Accurate prediction of bored pile settlement is a core element in reliable foundation design. While machine learning (ML) models demonstrate superior predictive capabilities compared to traditional analytical methods, their black-box nature and tendency to generate non-physical trajectories significantly hinder their practical application in geotechnical engineering. This study proposes a new computational framework that ensures physical consistency by combining Weighted Ensemble learning with SHAP (SHapley Additive exPlanations) interpretation capabilities. The integrated structure incorporates linear regression (LR), Random Forest (RF), XGBoost, and multilayer neural networks (MLP) to capture the complex interaction between piles and soil through a database of 563 field static load test records. A two-stage hybrid optimization strategy, using Random Search combined with a Genetic Algorithm (GA), was implemented to optimize the global weight set. The results showed that the GA-Ensemble model achieved excellent accuracy on the independent test set (R2 = 0.958, NRMSE = 5.27%, RMSE = 0.380, MAE = 0.2322), successfully neutralizing the ladder errors of the tree model to recreate a smooth, continuous, and mechanically compliant load-settlement (P-S) trajectory. Notably, by converting the SHAP value to a real physical scale (millimeters), the study clarified the algorithm’s inference logic. SHAP dependency analysis demonstrated that the model successfully captures the coupling characteristics of elastic compression and geometric nonlinearity by identifying the interaction between external force (P) and free length (Lf). Local diagnostic decompositions precisely extract failure mechanisms, successfully quantifying the “systemic penalty” caused by elastic shortening of structures and foundation puncture even in very stiff soil layers. In short, the proposed method has gone beyond the limits of a purely interpolation tool, establishing a transparent and geomechanically reliable expert system for predicting settlement.
SN - 1932-6203
ER -