Study of the optimal parameters on the stability of a large- height road embankment

Abstract

This thesis addresses critical challenges in slope stability assessment for high road embankments (6-

30m) through a dual-methodology framework integrating hybrid machine learning and physics- informed neural networks (PINNs). A comprehensive database of 1,176 finite element simulations

was generated using GeoStudio SLOPE/W, incorporating diverse soil properties and multi-berm geometries representative of Algerian mountainous highway networks. Field validation was conducted on 20 real embankments from Tebessa region to ensure practical applicability. The hybrid stacking ensemble—combining XGBoost, Support Vector Regression, Multi-Layer Perceptron, and Random Forest—achieved exceptional performance (R2=0.9978, RMSE=0.0199), delivering ~85% computational efficiency gains versus traditional FEM while maintaining high accuracy. SHAP analysis revealed three physically significant parameters as primary stability drivers: cohesion (47.7%)—representing the soil's intrinsic shear strength and resistance to sliding along potential failure surfaces; friction angle (31.1%)—governing the mobilization of normal stress into shear resistance, critical for stress redistribution under loading; and embankment height (6.6%)— controlling gravitational driving forces and stress state intensity within the slope mass. This hierarchy confirms alignment with geotechnical principles and the fundamental mechanics of limit equilibrium. The novel PINN framework explicitly embeds Mohr-Coulomb failure criterion, Bishop's Simplified Method, and monotonicity constraints into neural network training through composite loss functions optimized via Bayesian optimization. This approach achieved R2=0.9787 with perfect monotonicity compliance (100%) for critical parameters and overall physics satisfaction of 88.7%. Monte Carlo dropout provided well-calibrated uncertainty quantification (94.6% prediction interval coverage), enabling risk-informed decision-making. The validation dataset from Tebessa region demonstrated exceptional predictive accuracy: hybrid ensemble achieved +0.82% mean error, while PINN recorded +0.89%—both within acceptable engineering tolerances. The frameworks deliver real-time inference (0.8-1.2ms per prediction), representing >10,000× speedup versus traditional methods. The validated frameworks enable preliminary stability screening, parametric design optimization, and probabilistic risk assessment for infrastructure projects. An integrated three-stage workflow leverages hybrid ensemble for rapid screening, PINN for physics verification, and selective FEM for critical cases, reducing overall analysis time by 85-95% while maintaining rigorous safety assurance.