Machine learning aided stochastic reliability analysis of spatially variable slopes

Abstract This paper presents machine learning aided stochastic reliability analysis of spatially variable slopes, which significantly reduces the computational efforts and gives a complete statistical description of the factor of safety with promising accuracy compared with traditional methods. Within this framework, a small number of traditional random finite-element simulations are conducted. The samples of the random fields and the calculated factor of safety are, respectively, treated as training input and output data, and are fed into machine learning algorithms to find mathematical models to replace finite-element simulations. Two powerful machine learning algorithms used are the neural networks and the support-vector regression with their associated learning strategies. Several slopes are examined including stratified slopes with 3 or 4 layers described by 4 or 6 random fields. It is found that with 200 to 300 finite-element simulations (finished in about 5 ~ 8 h), the machine-learning generated model can predict the factor of safety accurately, and a stochastic analysis of 105 samples takes several minutes. However, the same traditional analysis would require hundreds of days of computation.