Probabilistic analysis of laterally loaded piles using response surface and neural network approaches

The response surface and the neural network methodologies are two approaches that are commonly used in reliability analysis of geotechnical problems with implicit performance functions, to deal with the complexity of probabilistic analyses. This paper proposes a two-step hybrid approach for reliability analysis. The first step obtains the design point using the first-degree polynomial response surface model. The second step constructs a neural network model of the performance function at the design point. The proposed method is first illustrated for a hypothetical laterally-loaded pile with analytical solutions. The case of a laterally-loaded steel pipe pile in Arkansas River sand is then presented, which involves non-normal random variables and spatial autocorrelation of soil strength parameters. Comparisons are made with Monte Carlo simulations incorporating importance sampling. Reliability-based parametric studies are performed on the Arkansas River example using the proposed hybrid approach. The influences on the reliability index and the probability of failure by the lateral load, depth of water table and correlation coefficient between unit weight and friction angle are investigated and discussed.