AI-driven predictions of the dynamic properties of fine-grained soils in South Italy based on laboratory testing

Abstract

This study presents the use of Artificial Intelligence (AI) to predict the dynamic behaviour of fine-grained soils of South Italy based on a detailed laboratory investigation. The investigation consists of Resonant Column (RC), Cyclic Torsional Shear (CTS), and Cyclic Triaxial (CTx) tests performed on 25 specimens of fine-grained soils retrieved from 11 sites in Sicily (South Italy). To develop accurate predictive models of soil dynamic properties, essential for site response analyses and dynamic soil-structure interaction, various regression techniques were applied. These techniques range from Multiple Linear Regression (MLR) to more complex AI methods, specifically Machine Learning (ML) and Deep Learning (DL) based on FeedForward Neural networks (FFN). Three predictive models were developed to derive strain-dependent shear modulus (G), damping ratio (D), and normalized shear modulus (G/G₀), using four inputs: shear strain (γ), plasticity index (PI), confining pressure (p’₀), and the Over Consolidation Ratio (OCR). To determine the optimal FFN topology, 1350 networks were developed by varying hidden layers (1–3), hidden neurons (1–50 per layer), and activation functions (ReLU, sigmoid and hyperbolic tangent). Hybrid FFN optimised through Genetic Algorithm and Particle Swarm Optimization techniques were also investigated. Single-hidden layer networks with fewer than 15 neurons provided acceptable predictions (R²_test of 0.97 for G-γ, 0.93 for G/G₀-γ, and 0.85 for D-γ models). Multiple-hidden layer networks yielded higher accuracy for G and D models but are more complex for practical use. The FFN models outperformed MLR and other established empirical formulations, highlighting the site-specificity of the modelling parameters of the latter.