From lab to algorithm: a review of AI-driven predictive modeling in soil stabilization for road construction

Abstract Soil stabilization for road construction relies on resource-intensive empirical methods, necessitating advanced data-driven solutions. This systematic literature review (SLR) and bibliometric analysis of 180 peer-reviewed studies (2015–2025) evaluates artificial intelligence (AI) applications in predicting and optimizing stabilized soil properties. Machine learning (ML) models, notably ensemble methods like Random Forest and Gradient Boosting Machines, consistently outperform traditional approaches, achieving R 2 > 0.90 and RMSE reductions up to 50% for critical geotechnical parameters such as Unconfined Compressive Strength (UCS) and California Bearing Ratio (CBR). Bibliometric trends indicate exponential research growth post-2019, evolving from accuracy-driven models (e.g., ANNs, SVMs) to interpretable frameworks employing SHAP and LIME, which identify binder content (e.g., cement, lime, GGBS) as the primary UCS driver (explaining up to 90% variance). Despite high predictive fidelity, challenges include a lab-to-field performance gap due to unmodeled environmental variability and the interpretability paradox of opaque high-accuracy models, risking misleading insights. Future research must prioritize comprehensive datasets linking laboratory and field performance, physics-informed ML to enhance generalization, and multi-objective optimization frameworks to balance mechanical strength, cost, and sustainability (e.g., reducing CO₂ emissions via industrial by-products like fly ash). This review positions AI as a transformative tool for designing resilient, eco-friendly road infrastructure, offering a roadmap for addressing current limitations and advancing geotechnical engineering practice.