AI-driven prediction of shear strength parameters in clay soils using experimental data

Abstract

Abstract In the context of post-earthquake reconstruction, the mechanical characterization of soils has become a crucial challenge to ensure the stability of structures in sensitive areas. Using artificial neural network (ANN) models to predict these mechanical properties presents an efficient and reliable approach. This study proposes the development and validation of an ANN model for predicting direct shear parameters, specifically cohesion and the friction angle, of clay soils from an earthquake-affected region in Morocco. A controlled laboratory campaign was conducted on clay soil collected from the Al Haouz region in Morocco. The campaign generated 35 distinct compaction configurations. For each configuration, four direct shear specimens were tested to determine the corresponding cohesion and friction angle, yielding an experimental basis of 140 direct shear specimens. The ANN model was then developed using the 35 input–output configurations derived from these tests, with 30 configurations used as development data and 5 additional configurations reserved as independent hold-out cases for final testing. In addition, a 5-fold cross-validation procedure was performed on the 30 development configurations to provide a more robust internal assessment of model stability. The ANN model included three input parameters (compaction energy, optimum water content, and dry density) and three hidden layers. The developed ANN predictive model showed close agreement between predicted and measured values for the five independent hold-out configurations, achieving coefficients of determination of 0.931 for cohesion and 0.967 for the friction angle. These results highlight the potential of the proposed ANN framework as a rapid predictive support tool for geotechnical assessment under conditions close to those investigated in this study, particularly in post-disaster contexts.