Few-shot learning-aided constitutive modelling of granular soils with knowledge transfer across different soil types

This study conducts the knowledge transfer among different soil types by adapting the pre-trained Leighton Buzzard Sand (LBS)-based long short-term memory (LSTM) model from the LBS database to the Fujian River Sand (FJS) database. After carefully discussing the continuous learning ability, a branch-based continuous learning strategy was proposed to accurately predict among different soil types. The main contributions are as follows: (1) The generated LBS database contains 464 micro-tomography (μCT)-based discrete element method (DEM) simulations of triaxial tests, while the FJS database contains 28 μCT-based DEM simulations; (2) Based on the LBS database, a dynamic cyclic training strategy was first proposed to predict both the deviatoric stress and volumetric strain points precisely; (3) A few-shot learning strategy was applied to implement the knowledge transfer across different soil types with excellent performance in both training and testing datasets; and (4) A branch-based continuous learning strategy was proposed with the LSTM layers to learn the basic variation trends of granular soils and the Fully Connected (FC) layers on different branches to adapt the LSTM layer mid-results to different soil types. The experimental results of different soil types with excellent predictive performance further validated the proposed branch-based continuous learning strategy.