<p>Accurate calibration of mesoscopic contact parameters in granular materials is essential for elucidating their multiscale mechanical behaviors and enhancing the design of geotechnical structures. This study proposes a novel self-adaptive framework integrating Box-Behnken Design (BBD), Non-dominated Sorting Genetic Algorithm II (NSGA-II), and DEM simulations for efficient parameter inversion and particle shape characterization. By replacing conventional DEM-intensive iterations with surrogate regression models embedded with a dynamic self-correction mechanism, the proposed method achieves an 81% reduction in computational demands while maintaining inversion accuracy within 5%. Furthermore, the box-counting dimension was introduced to quantitatively reveal the fractal characteristics of particle morphology and its profound impact on fabric anisotropy, strength evolution, and shear dilation behaviors. It is demonstrated that spherical particle replacement enhances peak strength, mitigates dilatancy, and facilitates stress transmission by reorganizing the contact network, leading to a preferential alignment of contact normals with the principal stress direction. This study bridges the micro–macro mechanical response of granular materials via a robust multiscale modeling paradigm, providing a transferable methodology for optimizing granular material design in complex geotechnical applications.</p>