Machine Learning-Based Prediction of the Seismic Bearing Capacity of a Shallow Strip Footing over a Void in Heterogeneous Soils

Abstract

The seismic bearing capacity of a shallow strip footing above a void displays a complex dependence on several characteristics, linked to geometric problems and to the soil properties. Hence, setting analytical models to estimate such bearing capacity is extremely challenging. In this work, machine learning (ML) techniques have been employed to predict the seismic bearing capacity of a shallow strip footing located over a single unsupported rectangular void in heterogeneous soil. A dataset consisting of 38,000 finite element limit analysis simulations has been created, and the mean value between the upper and lower bounds of the bearing capacity has been computed at the varying undrained shear strength and internal friction angle of the soil, horizontal earthquake accelerations, and position, shape, and size of the void. Three machine learning techniques have been adopted to learn the link between the aforementioned parameters and the bearing capacity: multilayer perceptron neural networks; a group method of data handling; and a combined adaptive-network-based fuzzy inference system and particle swarm optimization. The performances of these ML techniques have been compared with each other, in terms of the following statistical performance indices: coefficient of determination (R2); root mean square error (RMSE); mean absolute percentage error; scatter index; and standard bias. Results have shown that all the ML techniques perform well, though the multilayer perceptron has a slightly superior accuracy featuring noteworthy results (R2= 0.9955 and RMSE= 0.0158).