Estimation of seepage velocity and piping resistance of fiber-reinforced soil by using artificial neural network-based approach

Seepage erosion is one of the most important parameters leading to the failure of hydraulic structures like embankments or dams. The present study therefore focuses on developing artificial neural network-based models for preliminary prediction of seepage velocity and piping resistance of fiber-reinforced soil. A wide range of input data with variations in fiber type, soil type, fiber length, and fiber content is taken into account for developing artificial neural network (ANN) models. The performance of ANN models is also compared to that of popular regression models. Moreover, the sensitivity analyses based on Garson's algorithm and connection weight approach are conducted to rank effective input parameters in ANN formulas. The results indicate that ANN models show great performance in predicting these problems with high coefficient of determination (R-2 = 0.995 for seepage velocity and R-2 = 0.998 for piping resistance). ANN models are more accurate than regression models in predicting seepage velocity and piping resistance. For example, the root-mean-square error of ANN model for seepage velocity (0.013 cm/s) is significantly smaller than that of the quadratic regression model (0.099 cm/s). According to the connection weight approach, the top three effective parameters on seepage velocity are hydraulic gradient, gravel-sand, and silt-clay, while gravel-sand, critical hydraulic gradient, and specific gravity of soil are three parameters affecting piping resistance the most. In summary, the reliability and precision of ANN models for the estimation of seepage velocity and piping resistance are authenticated.

Automated summary

This study focuses on developing artificial neural network-based models for preliminary prediction of seepage velocity and piping resistance of fiber-reinforced soil. The results indicate that ANN models show great performance and accuracy in predicting these problems.