Machine Learning-Based Prediction of Shear Modulus Variation in Marine Clay

Résumé fourni par la source

Abstract The normalized shear modulus reduction curve (G/Gmax versus cyclic shear strain) is a critical parameter in the dynamic soil characterization for offshore seismic site response analyses. When dynamic laboratory testing is limited, the G/Gmax curve has been predicted with traditional approaches—semi-empirical correlations derived from limited datasets—often fail to capture the full variability of the G/Gmax curve of marine clays specially at the transition between medium to high cyclic shear strains, and high strains. This is where artificial intelligence (AI)—and more specifically, Deep Learning Algorithm (DLA)—offers transformative potential. AI has already demonstrated success in various civil engineering applications, from structural health monitoring to construction automation. However, its application to predicting the nonlinear dynamic behavior of soils remains largely unexplored. This study introduces a Deep Learning Approach, to predict the G/Gmax curve for Bay of Campeche clays with carbonate content ranging from 10% to 94%. A comprehensive database of 125 modulus reduction curves was compiled from resonant column and strain-controlled cyclic direct simple shear tests. The DLA model was implemented using a fully connected artificial network within the TensorFlow framework, incorporating shear strains and key geotechnical parameters such as in situ confining stress, overconsolidation ratio (OCR), void ratio, and liquidity index. Model training employed standard protocols, including early stopping and feature scaling, to ensure robust generalization. Performance evaluation against three widely used semi-empirical correlations—Vucetic & Dobry (1991), Darendeli (2001) and Taboada, et al., (2017)—demonstrated the superior predictive capability of the DLA, achieving a coefficient of correlation R2 of 0.95 and maintaining accuracy across both small-strain and nonlinear ranges. Unlike traditional models, the DLA effectively captured complex soil behavior without predefined functional assumptions, offering adaptability for diverse soil conditions. These findings highlight the potential of deep learning to improve G/Gmax curve prediction to enhance seismic response modeling and support safer, cost-effective design of structures when dynamic laboratory testing is limited or cannot be performed across all soil layers at the site.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.