A nonparametric regression algorithm, called the multivariate adaptive regression splines(MARS), is adopted to establish a simple and interpretable model to approximate the nonlinear interactions between inputs and outputs. Based on well-documented case data, MARS is capable of producing explicit expressions using the combination of simple linear spline functions, as well as obtaining the relative importance of input parameters. The reliability and accuracy of MARS were validated via the geotechnical applications of pile drivability assessment and stability evaluation of underground caverns. The coefficients of determination for the testing set in the two examples are 0. 921 and 0. 986, respectively, indicating that MARS can effectively fit the correlation between input parameters and target outputs. As for the parametric relative importance, the elasticity modulus of pile and quality of surrounding rock are recognized as the two most influential factors. MARS method can accurately relate the target responses to nonlinear and multivariate inputs; thus, it is useful in the practical engineering construction.