Ensemble learning for predicting the damped response of granular materials and interpretation through feature...

This study explores the application of ensemble learning, decision tree, random forest, and XGBoost, to predict the damping response of irregular granular materials using a data-driven approach. Training and testing data were measured using an acrylic beam excited with a low-amplitude sinusoidal sweep waveform (root mean square <1 N). Six granular materials, featuring both spherical and irregular particle shapes and varying mechanical properties, were selected as infill for the beam cavity. The main aim was to map between six input features: force, filling ratio, granular materials properties, and two targeted outputs: reduction of the beam's first resonance peak and root mean square acceleration. Feature importance analysis was conducted for each model developed to assess the dependence of the target values on the input features. It is concluded that ensemble models can offer accurate estimates (R2 ≥ 90%) with the XGBoost model incorporating all input features in the estimation. An attempt to estimate the damped response of hard-soft granular mixtures using empirically determined mixture-equivalent properties gave good estimates that differ from simple mass/volume-fraction averaging, with performance varying by filling ratio and excitation amplitude.