Experimental Investigation and ANN-Based Prediction of Dynamic Slope Responses: Crest Settlement, Toe Displacement and Response Time

Abstract

Dynamic behavior of soil slopes is controlled by the complex interaction of geometric conditions, characteristics of external loads, and reinforcement interventions. In most cases, the nonlinear interactions between the response variables and the effect of the multiple parameters including slope height, inclination, excitation frequency, amplitude and reinforcement condition, cannot be captured by the conventional empirical methods. In the present study an experimental database of 159 observations was consolidated to examine the dynamic stability behavior of reinforced and unreinforced slopes. The experimental program consisted of the slope heights of 300, 400 and 500 mm, slope inclinations of 30° to 40°, excitation frequencies of 1.0–2.0 Hz, amplitudes of 15–30 mm and reinforced and unreinforced conditions. The primary output responses considered were crest settlement, toe displacement, and response time. INDIVIDUAL Artificial Neural Network (ANN) models were developed for crest settlement, toe displacement and response time prediction. The model development was done with five different input parameters i.e., slope height, slope inclination, reinforcement condition, excitation frequency and loading amplitude, whereas the oscillation cycles were not considered as they were kept constant at 10 throughout the experimental program. The dataset was divided into 127 development observations and 32 independent test observations and five-fold cross-validation was carried out during model development to minimize potential information leakage between related experimental conditions.  The ANN model had the best prediction performance for crest settlement with the coefficient of determination of independent-test R2=0.943, RMSE=16.55mm and MAE=11.35mm. Also, good predictions were found for toe displacement (R2 = 0.866, RMSE = 35.63 mm, MAE = 30.11 mm), while response time showed moderate predictability (R2 = 0.708, RMSE = 0.419 s, MAE = 0.342 s).  The results show that the response-specific ANN framework is capable of reproducing important non-linear trends in the experimental dataset, with varying predictive performance across the three response variables. Response surface analysis based on ANN was used to further illustrate the combined effect of excitation frequency and amplitude, and reinforcement-effect surfaces were interpreted strictly within the domain of observations made experimentally. The study thus provides a framework for data-driven prediction and interpretation of dynamic slope responses informed by experiments, while acknowledging the limitations of dataset size, experimental domain and response-specific uncertainty.