Identifying Soft Soils using Pore-Pressure Parameters: A Machine Learning Approach

Abstract

- The cone penetration test (CPT) is a widely used method for identifying soil profiles and estimating soil parameters. Numerous correlations have been established to facilitate geo-characterization of soils based on CPT data. However, caution must be exercised when applying these correlations and laboratory tests should be used to validate them. Tropical residual soils are highly variable, even for seemingly similar samples, which can make it difficult for project designers to accurately characterize them. The present work focuses on a case study where the goal was to distinguish and characterize two soft soils existent on the foundation of a tailings dam in the southwest of Brazil. The construction of the dam is still ongoing, and its foundation belongs to a complex geological environment with soft soils that can reach N SPT blows as low as its own weight. The geological survey identifies two horizons of residual soil of dolomitic phyllite: soft and very soft. However, spatially distinguishing this material regarding its consistence has shown to be a challenging task. Since they differ essentially on the degree of weathering, most parameters for both materials are quite similar, and from laboratory tests, the parameter that helps differentiate these soils is the pore pressure Skempton parameter at failure – A f . Based on these findings, it can be inferred that the pore-pressure parameter Bq in CPT represents the excess pore-pressure during the test, whereas Af describes the excess pore-pressure at failure during triaxial tests. Despite the lack of a currently established theoretical correlation between the two parameters, they can offer valuable insight into the soil's response to rapid loading. Notably, both measures have proven to be effective in distinguishing between residual soils, even though they are distinct measures. In this study, the B q and A f parameters are employed to classify soils using an unsupervised learning method, specifically the K-means algorithm. The resulting clusters exhibit strong agreement with borehole profiles near the CPT locations.