Predicting Rutting Deformation in Flexible Pavements of Southwest Nigeria: A Machine Learning-Based Surrogate Model Derived from 3D Finite Element Analysis
1Apara, O. O. 1Akintayo, F. O.
1Department of Civil Engineering, Faculty of Technology, University of Ibadan, Ibadan, Oyo State, Nigeria
DOI: 10.36108/laujoces/6202.71.0140
Abstract
Rutting deformation in flexible pavements of Southwest Nigeria constitutes a persistent infrastructure failure. Existing CBR-based design methods do not adequately capture the multi-variable interaction between subgrade stiffness, pavement thickness, and traffic loading, while three-dimensional finite element analysis remains inaccessible to practising engineers due to computational cost. This study developed a machine learning-based surrogate model trained on 500 PLAXIS 3D finite element simulations generated via stratified Latin Hypercube Sampling across four USCS-classified subgrade groups: Lagos Blue Clay (CH/MH), Soft Laterite (CL/SC), Medium Laterite (SC/GC), and Stiff Laterite (GW/GC). The PLAXIS 3D Hardening Soil model was validated against the Burmister-Boussinesq elastic solution, which underestimated deformation by a factor of 5.2 at the stiffest subgrade tested, indicating the necessity of the nonlinear constitutive model. Three machine learning algorithms were trained and compared: XGBoost, Random Forest, and an Artificial Neural Network. The XGBoost surrogate achieved cross-validation R² = 0.9605 ± 0.0042, RMSE = 0.4598 mm, and MAPE = 11.02%, meeting all three target thresholds (R² > 0.90; RMSE < 0.50 mm; MAPE < 15%). Friction angle and subgrade stiffness emerged as the dominant predictor variables at 29.24% and 28.13% of feature importance respectively. Independent validation against five PLAXIS 3D simulations yielded an overall mean absolute percentage error (APE) of 13.9%. Excluding the single extrapolation case outside the training domain, the four interpolation cases achieved a mean APE of 5.88%, demonstrating excellent predictive performance within the calibrated parameter space for interpolation cases. The surrogate model output is structured as an architectural interface to the FERMA Pavement Condition Index framework, with full Architectural interface to FERMA PCI framework requiring a transfer function relating single-load Uz to cumulative rut depth as future work.
Keywords: Rutting deformation; Flexible pavements; Finite element analysis; Machine learning; XGBoost; Surrogate modelling; Lateritic soils; Pavement performance prediction; Southwest Nigeria.