Soil-Compressibility Prediction Models Using Machine Learning

The magnitude of the overall settlement depends on several variables such as the compression index, Cc, and recompression index, Cr, which are determined by a consolidation test; however, the test is time consuming and labor intensive. Correlations have been developed to approximate these compressibility indexes. In this study, a data driven approach has been employed to estimate Cc and Cr. Support vector machines classification is used to determine the number of distinct models to be developed. Classification accuracy is used for detecting the existence of separability between different soil classes that in turn is indicative of the number of models needed. The statistical models are built through a forward selection stepwise regression procedure. Seven variables were used, including the moisture content (w), initial void ratio (eo), dry unit weight (γdry), wet unit weight (γwet), automatic hammer SPT blow count (N), overburden stress (σ), and fines content (-200). The results confirm the need for separate models for three out of four soil types, these being coarse grained, fine grained, and organic peat. The models for each classification have varying degrees of accuracy. The model for the fine grained classification performs on par with existing correlations, with respect to Cc, whereas the models for coarse grained and organic peat classifications perform considerably better than that of existing correlations. The models generated also incorporate several factors not utilized in correlations from previous literature. These factors include the fines content (-200), automatic hammer blow count (N), and the interactions between the wet and dry density (γwet and γdry)