Machine learning regression approach for analysis of bearing capacity of conical foundations in heterogenous and anisotropic clays

An upper bound (UB) and lower bound (LB) finite element limit analysis cooperating with a machine learning method is adopted as a new solution for predicting the bearing capacity of conical foundations embedded in anisotropic and heterogenous clays. The anisotropic and heterogenous clays are simulated by anisotropic undrained strength (AUS) model for capturing the anisotropic strengths of clays. The bearing capacity of the conical foundation is investigated using the dimensionless parameter approach. The bearing capacity factors, as well as the failure mechanisms of conical foundations, are examined through 1296 numerical cases with changing of four input dimensionless parameters, namely cone apex angle, embedded depth ratio, the anisotropic ratio, and the strength gradient ratio. Based on numerical results, a machine learning technique of multivariate adaptive regression splines (MARS) model is used for accessing the sensitivity of each investigated dimensionless parameter and functioning the relationship between input parameters and output bearing capacity factors. The results of the analysis are prepared in charts, design tables, and empirical equations from MARS. The paper can be the theory guidelines for initial design and provide an effective tool for practitioners in determining the bearing capacity of conical foundation embedded in anisotropic and heterogenous clays.