Abstract:
Understanding soil behaviours are fundamental to geotechnical design. Myriad empirical and analytical models have been proposed for prediction accordingly but they tend to be site-specific and increasing parameters need to be calibrated for constitutive models. With the increasing data in the geotechnical domain, machine learning (ML) has emerged as a new methodology to directly learn from raw data to identify soil behaviours. Its applicability has been proven to be promising because of its versatility and strong fitting capability. Nevertheless, the current ML-based data-driven models still exhibited limitations including lack of interpretability, dependency on numerous high-quality data and poor generalization ability, thus they are still far away from application to engineering practice. To this end, this study aims to elaborate on data-driven models for predicting soil mechanical behaviours and facilitating their applications in engineering practice. A multi-fidelity residual neural network incorporating physical constraints and Monte Carlo Dropout uncertainty is proposed to leverage existing knowledge and limited high-quality data for modelling the mechanical behaviours of soils. The developed data-driven model is also integrated with finite element code for modelling boundary value problems compared with conventional numerical modelling methods. The results indicate the developed data-driven model successfully captures monotonic, cyclic and rate-dependent of soil behaviours. The modelling framework is generalizable and generic, showing a large potential to be applied to model behaviours of various materials.
Dr Zhang will also to briefly introduce his Royal Society fellowship project, titled 'Deep Learning-Based Multi-Scale and Multi-Physics Models for Granular Materials'.