Machine learning techniques for relating liquid limit obtained by Casagrande cup and fall cone test in low-medium plasticity fine grained soils

Abstract

The liquid limit is a key property of fine soils closely related to the stress-strain behaviour and other relevant characteristics of soils such as the expansive potential. There are two standardized methods for its determination, the Casagrande cup and the fall cone test, among which there are many correlations that offer heterogeneous results. In the present study, a compilation of 113 data from fine soil samples with low-medium plasticity has been carried out. Then, a comparative study of different machine learning algorithms was carried out to relate the liquid limit obtained from both methods, having in consideration other parameters such as the plastic limit and the percentages of passing through the 0.40 and 0.075 mm sieves. The result of this study has shown that Extremely randomized trees algorithm provides the best performance. Consequently, the algorithm has been tuned to enhance the precision, obtaining a coefficient of determination (R2) value of 0.99. The results demonstrate the potential of machine learning techniques for relating liquid limit obtained by Casagrande's method and fall cone test in fine-grained soils with low-medium plasticity, mainly for values of the liquid limit higher than 30 for which classical linear regression approaches provide lower performance metrics.