Predicting the cyclic behavior and strength-weakening effect of saturated clays using artificial neural network

The cyclic behavior of soft cohesive soils under shear loads is characterized by progressively increasing strain and by the growth of the pore pressure that can lead to an effective stress reduction and, eventually, to the sudden failure of the soil, both risks with evident engineering safety implications. Although this problem has received much attention in research, most present approaches can only predict some parameters of the clay performance separately, commonly leading to highly complex approaches requiring extensive training and expertise. The present research uses an Artificial Neural Network (ANN) and machine learning to predict the cyclic behavior of the soft clays in the investigated site, considering for the first time all the relevant parameters that characterize the problem. The proposed ANN includes 9 inputs, two hidden layers with 10 neurons each, and five outputs. Nine inputs include the vertical effective consolidation pressure, parameters from the monotonic shear test, and defined input variables from the cyclic simple shear test. As outputs, the net considers five different results, including the various parameters of the shear strain response for each cycle, the maximum number of cycles, and the pore pressure increase. The resulting ANN shows predictions with high accuracy, with R = 0.995, and individual errors below 10 % in most cases. No prior training or experience is required to use the ANN, and it can be confidently used as an alternative to other analytical and numerical approaches for analyzing clay cyclic behavior, as long as the input values fall within the defined ranges.