Data-driven generative adversarial network and machine learning hybrid framework for predicting rutting depth in recycled asphalt shingle pavements

Excessive rutting is a leading cause of premature asphalt pavement deterioration, resulting in reduced service life and safety. Consequently, growing interest in sustainable materials has driven research into recycled asphalt shingles (RAS), which contain a highly viscous binder that boosts rutting resistance and lowers costs by reducing binder demand and minimising shingle waste. However, limited and costly testing has restricted research, emphasising the need for machine learning (ML) models to predict the rutting depth (RD) in RAS-asphalt mixtures. Yet, progress in ML deployment has been hampered by scarce and imbalanced data, limiting prediciton reliability. Therefore, this study adopts conditional tabular generative adversarial networks to address data scarcity and produce a balanced dataset for training advanced ML models. This process enhanced the generalisation and effectiveness of the models trained on synthetic data; the XGB model achieved an R² of 0.9872 and a MAPE of 7.49%. Moreover, feature importance analysis indicated that the number of passes, mixture gradation, and bitumen content are the significant factors influencing RD inRAS-asphalt mixtures. This work shows the successful combination of synthetic data and advanced ML models to promote sustainable infrastructure. It also develops a user interface for estimating the RD, offering stakeholders an efficient and rapid alternative to costly experiments.