A Hybrid Mechanistic–Empirical and Neural Network Model Framework for Forecasting Fatigue Crack Deterioration in Ethiopian Flexible Pavements

Abstract

Fatigue crack deterioration in flexible pavements results from structural loading, traffic demand, material aging, and climatic exposure; yet, Ethiopian pavement models remain largely empirical, with little mechanistic foundation. This study develops a hybrid mechanistic–empirical and artificial neural network framework to forecast fatigue crack progression along a five-kilometer segment of the Woldia–Jeneto road in northern Ethiopia, built in 2015 and assessed after ten years of service. ERAPave layered elastic analysis computed critical horizontal tensile strain at the asphalt base, using the ERA manual recommendation of the Australian fatigue criterion for tropical areas, deriving cumulative damage indices via Miner’s rule. These outputs, alongside material properties, soil indices, traffic, climate, and temporal variables, formed an 18-feature input vector, which was trained using Latin–Hypercube Sampling and leave-one-out cross-validation under data-scarce conditions. Critical tandem-axle loads of 200.2 kN produced tensile strains of 182.7–199.83 με and damage ratios of 0.39–0.76 within fatigue lifetimes of 10.46–20.12 million ESALs, exceeding the 7.93 million ESAL design threshold. The model achieved R2 = 0.9997 and MAPE = 1.64%; these figures reflect five-station training conditions and synthetic augmentation rather than unconditional generalization accuracy. Ten-year forecasts place Station 5 at structural failure within three years, supporting evidence-based pavement maintenance planning.