MULTI-SOURCE DATA-DRIVEN CONSTITUTIVE MODEL FOR CLAYSTONE USING TRANSFER LEARNING: METHODOLOGY, ALGORITHM AND APPLICATION

In recent years, data-driven constitutive models, developed using machine learning approaches, have become increasingly popular for the constitutive modelling of geomaterials. However, this emerging paradigm faces a significant challenge that experimental data are usually scarce due to the expensive and time-consuming disadvantages, while synthetic data generated rely on existing theoretical model are difficult to reflect the true characteristics of geotechnical materials. In this study, a multi-source data-driven constitutive model for Callovo-Oxfordian (COx) claystone is developed by integrating high-fidelity and low-fidelity data through a transfer learning strategy to tackle the data scarcity issue. The high-fidelity data are collected from experimental results published in the literature, while the low-fidelity data are generated by a theoretical constitutive model. An implicit stress point integration algorithm is then developed for the proposed data-driven model and implemented in the Finite Element Method software ABAQUS via a user subroutine. Finally, numerical simulations are conducted to analyse underground excavation in argillaceous nuclear waste repositories, demonstrating the applicability of the developed model to practical boundary value problems. Numerical tests on representative volume elements show that, compared to the theoretical constitutive model, the prediction accuracy of the proposed data-driven model for experimental data is significantly improved. In summary, this study presents a novel methodology for developing multi-source data-driven constitutive models for geomaterials, offering a valuable tool for the stability assessment of nuclear waste disposal facilities.