Abstract: Data-driven approaches provide new ways of modelling the mechanical behaviour of granular geotechnical materials. Due to the time-consuming, labor-intensive and costly nature of physical testing on granular geotechnical materials, existing research largely relies on synthetic, artificially generated data for model training. However, synthetic data generated by algorithms or models often has low fidelity and struggles to reflect the complexity and diversity of granular geotechnical materials. Consequently, data-driven models constructed in this manner are rarely applicable to real-world problems. This paper proposes an innovative, multi-fidelity, data-driven approach based on sequential transfer learning for the intelligent prediction of the properties of granular geotechnical materials. This method uses a multi-fidelity data fusion strategy to improve model prediction performance gradually through transfer learning. First, a large amount of low-cost, low-fidelity data is generated using macro-scale constitutive models in order to create a base model with strong generalization capabilities. Secondly, micro-scale numerical experiments are introduced by FDEM to obtain medium-fidelity data considering particle geometry, serving as a transitional bridge from low- fidelity to high-fidelity. Finally, a small amount of high-fidelity physical test data is used to refine the model further, significantly improving its predictive accuracy. This sequential transfer learning process enables a gradual transition from low-fidelity simulations to high-fidelity experimental scenarios, enhancing the model as it progresses. Validation results show that the constructed model can reproduce the stress-strain response of granular geotechnical materials under different loading conditions. The model’s predictive accuracy and generalization capability both outperform those of models trained on single datasets, significantly reducing the reliance of data-driven models on extensive physical test data. This method provides valuable insights into constructing robust, low-cost, data-driven constitutive models based on sparse experimental data.