Pavement backcalculation using artificial neural networks

Pavement structural evaluation constitutes an essential component of management and maintenance routines. In this context, the precise determination of layer stiffness properties (elastic moduli) is fundamental for pavement backcalculation, enabling more accurate in situ assessments, supporting quality control in construction and rehabilitation, as well as an enhanced estimation of structural service life. Given the intrinsically nonlinear nature of the backcalculation problem, approaches based on Nonlinear Least Squares (NLS) minimization have been employed, utilizing optimization algorithms such as Gauss-Newton (GN) and Levenberg-Marquardt (LM), which demonstrate robustness in their application. However, it is observed in iterative methods the convergence speed of NLS optimization algorithms can be perceptibly influenced by the choice of initial estimates (seed moduli). In this context, Artificial Neural Networks (ANNs) represent a promising alternative, particularly due to their capacity to directly map deflections to moduli with notable computational speed and efficiency after training. Training a neural network necessitates a vast dataset, which can be synthetically generated using models such as the Finite Element Method. Therefore, a hybrid model is proposed wherein an ANN trained with data from a precise FEM model, can provide fast and reliable initial estimates for the moduli, potentially improving the efficiency of the iterative NLS optimization processes. The computational efficiency, robustness, and accuracy of this hybrid model will be rigorously evaluated in the context of asphalt pavement backcalculation.