Prediction of Normalized Shear Modulus and Damping Ratio for Granular Soils Over a Wide Strain Range Using Deep Neural Network Modelling

Abstract

Dynamic properties (i.e., shear modulus and damping ratio) of geomaterials play a vital role in civil engineering applications and are essential for reliable dynamic response analysis. This paper presents a novel approach for predicting the normalized shear modulus (G/Gmax) and damping ratio (D) of granular soils across a wide strain range using a Deep Neural Network (DNN) modeling strategy. Traditional methods for predicting these properties often rely on empirically derived relationships that may not capture the full complexity of granular soil behavior under varying strain conditions. A comprehensive dataset of shear modulus and damping ratio measurements from laboratory cyclic triaxial (CT) and resonant column (RC) tests conducted under various conditions is utilized. The dataset covers a wide range of strain levels, allowing for a more robust and versatile modeling approach. For predicting the G/Gmax and D of granular soils, a Deep Feed-Forward Neural Network (DFFNN) model was developed to learn the features from input data. The proposed model considers the influence of grading characteristics (Gravel Content, GC, median particle size, D50, Uniformity Coefficient, Cu, and Coefficient of Curvature, Cc), shear strain (\(\gamma\)), void ratio (e), mean effective confining pressure (\({\sigma ^{\prime}_m}\)), consolidation stress ratio (KC) and specimens’ preparation method (S-P) as input data. The empirical models (EMs) and three other intelligent techniques, namely Shallow Neural Network (SNN), Support Vector Regression (SVR), and Gradient Boosting Regression (GBR) were used for comparison. The testing accuracy of the proposed DFFNN for predicting the G/Gmax and D was 0.9830 and 0.9396, respectively. The results demonstrate that the proposed DFFNN modeling strategy provides a highly accurate means of predicting G/Gmax and D for granular soils across a broad shear strain range. This method offers advantages over EMs by incorporating a data-driven approach that can adapt to the specific behavior of different granular soil types and loading conditions.