An improved multi-fidelity neural network for surrogate modeling of mechanical responses of irregularly shaped granular materials

High-resolution discrete element method (DEM) simulations are essential for accurately capturing the mechanical behavior of irregularly shaped granular materials; however, their high computational cost limits practical applications in large-scale engineering analyses. Data-driven surrogate models alleviate this burden, while multi-fidelity learning further enhances data efficiency by incorporating inexpensive low-fidelity (LF) simulations. Nevertheless, existing multi-fidelity approaches remain constrained by the limited parameter-space coverage of scarce high-fidelity (HF) samples. This study develops an improved multi-fidelity neural network (IMFNN) with a projection-based data enhancement strategy to predict the critical-state macro–mesoscopic responses of irregular granular materials. First, a dedicated multi-fidelity DEM dataset is established, where fidelity levels are defined by particle-shape resolution and controlled by the number of sub-spheres in the clumped-particle representations. A feedforward neural network is pretrained using abundant LF data to capture dominant nonlinear response trends. Subsequently, a projection neural network is trained using HF samples together with their LF-level predictions to learn the LF–HF correction relationship. The learned relationship is then propagated to LF samples to generate pseudo-HF labels, thereby extending the effective coverage of HF-level supervision. The real and pseudo-HF samples are integrated to train the final surrogate model, which enhances the utilization of multi-fidelity information and improves predictive performance. The IMFNN accurately predicts critical-state macroscopic and mesoscopic responses and outperforms a pooled feedforward neural network and a conventional multi-fidelity neural network under the considered HF-data availability levels. The trained IMFNN surrogate model further enables efficient parametric studies and probabilistic analyses, providing a computationally efficient alternative to repeated high-resolution DEM simulations of irregular granular materials.