ARTIFICIAL NEURAL NETWORKS IN SOIL SHEAR STRENGTH PREDICTION

Resumo

The estimation of soil shear strength parameters has been of much relevance in Geotechnical
Engineering since the early stages, leading to the creation of many correlations based on in situ tests.
However, Das and Basudhar (2008), Goktepe (2008) and Shooshpasha, Amiri and MolaAbasi (2014)
stated that these existing correlations have limited use and low generalization capacity when compared
to the neural models they proposed from index properties of soils. On behalf of that, this work was
carried on prediction of cohesion and friction angle of soils in the natural state by the use of
backpropagation multi-layer perceptrons built from easy-to-collected in situ input, NSPT, V0’ and soil
type. The architecture chosen, A:3-5-3-2, used hyperbolic tangent activation function, being trained and
tested by 38 soil samples, having reached correlations up to 0,94 for training, attesting its efficiency.