Development of Deep Learning Based Model for Estimating Erosion Rates in Cohesive Soils

The adverse effects of climate change have led to coastal communities facing increased frequency and intensity of extreme meteorological events such as storm surges. As storm surge intensifies, the main cause of concern for shallow foundation elements, common to residential construction in these communities, is scour and erosion. The flow behavior and the forces driving scour around foundations remain the same, independent of whether the soil is cohesive or cohesionless. However, with cohesive soils, an added source of resistance is found in the form of internal particle cohesion. As a result, the absence of direct testing makes estimating the erosion rates in the case of cohesive soils a difficult proposition. Therefore, this study was designed to develop a data-driven model to predict erosion rates for cohesive soils. The NCHRP database of erosion test results was used to develop a deep neural network model predicting erosion rates using soil index test results. The model showed strong predictive performance for low to moderate erosion rates. However, the lack of training samples with very high erosion rates led to the model underpredicting the erosion rates higher than 30 mm/min. Sensitivity analysis on the model performance demonstrated a positive correlation of erosion rates with bed shear stresses and a negative correlation with undrained shear strength of the soil. In contrast, the influence of plasticity index, water content, and fines fraction of the soils were found to be non-monotonic. Taking into account, the resources required to obtain site-specific erosion characteristics for cohesive soils, the data-driven method presented herein can serve as a valuable tool in obtaining first-order approximation of erosion rates for cohesive soils, thus aiding community resiliency efforts and rapid decision-making.