Artificial neural network for stress-strain behavior of sandy soils: Knowledge based verification

In this paper, artificial neural networks (ANNs) are applied to model the stress-strain behavior of in situ sandy soils containing nonplastic fines. A main drawback of these types of models is discussed, i.e. an ANN based model gives no information how the model inputs affect the output. A systematic approach is therefore presented to acquire and verify the stored knowledge of a general ANN based constitutive soil model. Sensitivities of the output to corresponding inputs are defined mathematically. A sensitivity analysis is then performed to extract the dominant rules of the proposed model, which compare favorably with experimental observations.