Random forest based artificial intelligent model for predicting failure envelopes of caisson foundations in sand

To reduce the computational cost and improve the accuracy in predicting failure envelopes of caisson foundations, this study proposes an intelligent method using random forest (RF) based on data extended from experiments and calibrated numerical simulations. Two databases are built from the numerical results by coupled Lagrangian finite element method and smoothed particle hydrodynamics with a critical state based simple sand model (CLSPH-SIMSAND). The first database involves the failure envelopes of caisson foundations with various specifications for a given sand, and the second database includes two additional failure envelopes of caisson foundations in other granular soils. The relationship between the characteristic measures of failure envelope and sand properties as well as the specification of caisson foundation is trained by RF using the prepared databases. The results indicate the RF based model is able to accurately learn the failure mechanism of caisson foundation from the raw data. Once a RF based model that can accurately reproduce the failure envelopes of caisson foundations in a given sand is developed, it can be easily modified to predict the failure envelopes of caisson foundations in a random granular soil as long as one numerical result in such soil is added to the database. Therefore, the RF based model is much more convenient than the calibration of parameters used in the conventional analytical solutions and the computational cost is much less than the conventional numerical modelling methods.