A machine learning approach for the prediction of very thin wearing layers asphalt concretes volumetric properties and performance

In recent years, many researchers in the field of pavement engineering have worked with the aim of developing a model capable of predicting the mechanical behavior of a mixture starting from its composition's parameters. This has been done following two different approaches. The first involved the use of advanced constitutive laws based on the materials mechanics; the second, instead of being physically based, was data-driven. The present work belongs to this second context and aims to present, implement and apply a strategy to develop the optimal model for solving an assigned predictive problem. Specifically, a Machine Learning approach, a Feedforward Backpropagation Shallow Neural Network, was investigated. The objective was to correlate stiffness modulus, air voids and voids in the mineral aggregate to the mixture main composition's parameters identified in: bitumen content, particle size and a categorical variable distinguishing the bitumen type and production site. Since the maximum aggregate size is 10 mm, the sieves considered were of 10, 6.3, 2, 0.5 and 0.063-mm diameters. The present study focused on 92 variants of asphalt concretes for very thin road pavement wearing layers produced both in plant and in laboratory. Despite the wide variation ranges of each parameter considered, the optimal model returns fully satisfactory performance. The overall Pearson correlation coefficient is equal to 0.9490, also by virtue of the innovative algorithms implemented as k-fold Cross-Validation (CV) and Bayesian Optimization (BO). These algorithms have allowed on the one hand the improvement of the model's predictive performance making them more reliable and, on the other hand, the optimization of hyperparameters and architecture. The methodology developed can become an important reference in this field since it is independent from the specific predictive application. In this sense, it can help other researchers in the fine-tuning of neural models in the field of pavement engineering.