TITLE:
Gradient Boosting Surrogates for Limit-Equilibrium Factor of Safety of Mine Slopes with ICA Hyperparameter Tuning: A Duplicate-Free, Group-Aware Evaluation on a Compiled Dataset Relevant to Tarkwa, Ghana
AUTHORS:
Isaac Mensah Quansah, Festus Kunkyin-Saadaari, Larry Testamen Karikari, Charles Asiedu
KEYWORDS:
Slope Stability, Factor of Safety, Gradient Boosting, Imperialist Competitive Algorithm, Surrogate Modelling, Group-Aware Validation, Open-Pit Mine
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.11,
November
10,
2026
ABSTRACT: Accurate and rapid estimation of the Factor of Safety (FOS) of mine slopes supports geotechnical screening in open-pit mining. This study evaluates whether gradient boosting models can act as surrogates for Bishop’s Simplified Method limit-equilibrium (LEM) FOS using seven input parameters (Rock Mass Rating, unit weight, cohesion, friction angle, slope angle, slope height and pore water pressure ratio). Three regularised learners (XGBoost, LightGBM, CatBoost) and three Imperialist Competitive Algorithm (ICA)-tuned counterparts were compared. The compiled dataset contained 281 records, of which only 103 were unique and 100 had unique input vectors; the remaining records were exact duplicates, which, when split at random, leak between training and test sets. After removing duplicates, records were grouped into 62 slope families (identical unit weight, cohesion, friction angle and slope height) and all validation kept families intact. Performance was assessed by repeated (2 × 5-fold) group-aware nested cross-validation and a final untouched test set of 13 families (20 records), with family-level bootstrap confidence intervals. Because FOS is itself computed from the model predictors, the models are LEM surrogates and were not validated against observed slope behaviour. Under this protocol, nested-CV R2 was low: CatBoost R2 = 0.27 (95% CI [0.11, 0.40], RMSE = 0.34) and ICA-CatBoost R2 = 0.23; only these two models improved significantly on a mean-only baseline (RMSE 0.39). No pairwise difference between CatBoost and ICA-CatBoost, or between any ICA model and its base learner, was statistically supported, and the ICA overfitting penalty did not improve held-out error relative to an unpenalised search. At the FOS = 1.0 threshold, nearly all unstable slopes were predicted as stable. Friction angle was the highest-ranked predictor in all six models, but with wide uncertainty. The results indicate that, for this dataset, the seven predictors do not support reliable FOS surrogate modelling, and that high R2 values on the raw dataset are largely attributable to duplicate leakage.