Abstract: The damage of the cemented paste backfill is an important factor affecting the safety and stability of the mine. In order to explore the damage characteristics of the cemented paste backfill, the full state of gangue cemented paste backfill (GCPB) crack evolution was investigated by indoor experiments, machine learning hierarchical cluster analysis and numerical simulation. The results show that when the average link length energy of characteristic links of GCPB samples PC-5, PC-10, PC-15, PC-20 exceeded 35 ms·mV, 50 ms·mV, 60 ms·mV, 150 ms·mV, the macroscopic main crack surface of developed, and the main crack damage surface changed from tensile to shear damage surface with the increase of Portland Cement (PC) dosing. During the damage process, there was a negative linear relationship between the characteristic link subset scale Ω and the damage variable D. The value of Ω could characterize the degree of damage evolution of the cemented paste backfill. Before the stress reaches the peak stress and post-damage residual strength, the link length value eigenspace correlation length ξ showed a significant increase. This could be the precursor feature of the beginning of macroscopic damage and the end of the first damage of the cemented paste backfill. Through PFC numerical simulation, the crack type and variation pattern can be observed effectively, which is in good agreement with the clustering results.