Cross-physical-field mechanical analysis and data-driven multimodal data generation for fractured sandstone based on particle discrete element simulation

To investigate the mechanical evolution mechanism of fractured sandstone, this study employed discrete element simulations to construct numerical models with varying fracture parameters. Systematic analysis was conducted on their mechanical responses and energy evolution under uniaxial compression. A U-Net-based diffusion generative model was developed, utilizing simulation results to build a sample library for training and validation. This enabled cross-modal mapping between fracture parameters and multi-modal data, including crack propagation images and the sequential data of stress and energy versus strain. Key findings include: (1) The fracture inclination angle controls failure modes by modulating the stress intensity factor ratio, thereby governing energy evolution characteristics. (2) Fracture length reduces the effective load-bearing area and intensifies stress concentration at the tips, promoting mixed-mode crack development. (3) Under stress arch effects, specimens with end fractures exhibit pseudo-plastic behavior during stress drops phase, accompanied by periodic “blockage–release” energy fluctuations. (4) The multi-modal generation model achieved a mean square error of 0.013 for sequence data, a peak signal-to-noise ratio of 32.5 dB, and a structural similarity index of 0.852 for crack propagation images, demonstrating high generative accuracy and reliability. This study offers an intellectually rigorous and tool-supported approach for stability assessment in geotechnical engineering. • Mechanistic insight: Elucidated the cross-physical-field mechanisms governing sandstone's mechanical and energy evolution, establishing a fundamental link between fracture parameters and macroscopic behavior. • Methodological innovation: A novel framework for intelligent multi-modal data generation was established, which pioneers the use of a U-Net-conditioned diffusion model to synthesize coupled crack propagation images and the sequential data of stress and energy versus strain. • Practical translation: Developed an end-to-end intelligent generation system that translates mechanistic insights into an efficient digital tool for rock engineering stability assessment.