Efficient reliability analysis considering uncertainty in random field parameters: Trained neural networks as surrogate models

This paper presents an efficient reliability analysis framework, by using trained artificial neural networks (ANNs) as surrogate models, for geotechnical problems where the random field parameters like the mean and standard deviation are themselves uncertain. Random field theory has been extensively used to model soil uncertainty and spatial variability. However, due to limited availability of data, random field parameters can rarely be estimated accurately, often estimated in confidence intervals (uncertain parameters). Monte Carlo based reliability analysis is computationally extremely demanding because the function to map outcomes of random fields to structural response can only be calculated via numerical simulations. The authors have used trained ANNs as surrogate models in reliability analysis. However, these ANNs are specific for random fields with deterministic parameters. This paper presents a new framework in which trained ANN models are for random fields with variable parameters. A key component is the design of experiments - generating representative outcomes. In the prediction of the bearing capacity for strip footings, the efficiency and accuracy of this framework are successfully demonstrated. This framework is also efficient in reliability sensitivity studies. One main finding is that ignoring random field parameter uncertainty could lead to underestimated failure probability and hence unsafe design.

Automated summary

This paper presents an efficient reliability analysis framework using trained artificial neural networks as surrogate models for geotechnical problems with uncertain random field parameters. By generating representative outcomes through experimental design, the framework demonstrates high efficiency and accuracy in predicting bearing capacity.