Estimation of effective cohesion using artificial neural networks based on index soil properties: A Singapore case

This study presents a development of a multi-layer perceptron (MLP) model to spatially estimate and analyze the variability of effective cohesion for residual soils that are commonly associated with rainfall-induced slope failures in Singapore. A number of soil data were collected from the various construction sites, and a set of qualified Nanyang Technological University (NTU) data were utilized to determine a criterion for data selection. Four index properties (i.e., percentage of fines and coarse fractions, liquid and plastic limits) were used as training parameters to estimate the effective cohesion of residual soils from different geological formations. Ordinary kriging analyses were carried out to analyze the spatial distribution and variability of effective cohesion. As a result, the appropriate effective cohesions can be estimated using the MLP model with the incorporation of the selected values of measured effective cohesion as training data and four index soil properties as input data. In the combination of estimated and measured effective cohesions, the spatial analysis using Kriging method can clearly differentiate the variations in effective cohesion with respect to the different geological formations. •Residual soils are commonly associated with rainfall-induced slope failure in Singapore.•Effective cohesion of residual soils was estimated using a multi-layer perceptron model.•Available soil data were spatially collected for a multi-layer perceptron training.•% particle fractions, LL, and PL are input variables in the multi-layer perceptron model.•The multi-layer perceptron model generates variability of c’ for Singapore residual soils.