Desenvolvimento de modelos neurais e mapas geológico-geotécnicos para pavimentação no estado do Ceará

Abstract

The development of neural models and geological-geotechnical maps to assist pavement projects in the state of Ceará is of paramount importance in optimizing and enhancing pavement design accuracy. Currently, there is a clear need for more advanced and specific tools to address the geotechnical challenges. Through advancements in the use of artificial neural networks in recent times, it has been possible to optimize geological-geotechnical characterization studies of pavements, offering an innovative approach to understanding the geotechnical properties of materials used in the foundation and layers of pavements. Neural modeling, when there is availability of adequate quantity and quality of data, enables the prediction of a variety of soil geotechnical characteristics, potentially offering a more economical and complementary estimation option compared to the traditional acquisition of data solely through laboratory tests of CBR (California Bearing Ratio) or ISC (Index of Soil Compaction) and Resilience Modulus (MR) necessary for pavement design. The main objective of this dissertation was to develop neural models through the use of artificial intelligence and geological-geotechnical maps to be applied to pavement projects in the State of Ceará. To achieve this, databases containing geotechnical information on pavements in the State of Ceará were compiled. In total, the databases contain more than eight thousand data points covering soil physical properties, CBR, and AASHTO (American Association of State Highway and Transportation Officials) classification information. All this data originates from a project of the Federal University in partnership with the DNIT (National Department of Infrastructure and Transportation). Through the use of Artificial Neural Networks (ANNs), neural models were developed to predict CBR and MR values, with most of these models achieving correlation values above 0.7, suggesting a significant relationship between the variables used to estimate CBR and MR. The best model developed for CBR had a correlation value of 0.98 in validation, while for MR, the most representative model had a value of 0.86. This highlights the promising ease and applicability of neural models in predicting CBR and MR for empirical and empirical-mechanistic pavement design. Through geological-geotechnical maps, it was possible to obtain a view of the geotechnical properties and distribution of geological elements in the analyzed sections, allowing for a closer approach to the selection of materials and construction techniques.