Machine learning techniques for evaluation of permanent deformation responses from geogrid stabilized pavements

Permanent deformation reduction (a.k.a., rut resistance capacity) and stiffness improvement are two key features of geogrid stabilized pavements . Furthermore, geogrids also contribute to preserving the uniformity of the stiffness distribution over a wide area, proving the increased reliability provided by stabilization. In most common pavement design and evaluation methodologies, permanent deformation is an essential component for long-term pavement performance assessment. For example, the AASHTO (Association of State Highway Transportation Officials) R50 standard considers permanent deformation for the derivation of the traffic benefit ratio (TBR) or base course reduction (BCR) factor. Although full-scale accelerated pavement testing or in-service pavement testing is ideal for assessing permanent deformation responses , such testing may not be feasible to perform in a wide range of situations, including diverse subgrade types, climatic zones , and material types . An alternative is to use large-scale plate load testing for in-situ material characterization. Automated Plate Load Testing (APLT) is a field-based plate load testing system for applying dynamic loads and measuring permanent and resilient deformations. For this paper, APLTs were conducted to measure permanent deformations on several pavement sections consisting of different aggregate base course (ABC) thicknesses, ABC material types, multi-axial geogrids, and subgrade conditions. Several machine learning techniques , including Multiple Linear Regression Analysis (MLRA), Gene Expression Programming (GEP), Customized Non-linear Regression (CNR), Traditional Machine Learning (TML), and Artificial Neural Networks (ANN), were explored to develop prediction models for permanent deformation. Among the TML models, Extra Tree, XGBoost, and LightGBM demonstrated superior accuracy and robustness against overfitting. These models effectively captured the complex interactions between model parameters, making them suitable for evaluating geogrid-stabilized pavements.