Abstract: A two-stage XGBoost-based prediction framework was developed to evaluate the damage potential of mainshock-aftershock (MA) sequences. Reinforced concrete (RC) bare frame structures and RC frame structures with infilled walls were selected as the research subject. A total of 662 real mainshock-aftershock records were selected for nonlinear time history analyses. Hysteretic energy was adopted as the structural damage indicator. Four damage indices were established: mainshock hysteretic energy (E_H,MS), incremental aftershock hysteretic energy (ΔE_H,AS), accumulated hysteretic energy (E_H,MA), and accumulated hysteretic energy ratio of aftershock (γ). These indices corresponded to four intensity representations: mainshock intensity, aftershock intensity, combined MA intensity, and MA intensity ratio. Three critical intensity measures (IMs) were selected through the correlation analysis from 30 candidate IMs based on low information redundancy and on strong logarithmic correlations with structural damage. The XGBoost algorithm was first employed to predict mainshock damage potential using the identified critical IMs. Subsequently, the predicted mainshock damage potential and aftershock IMs were utilized as key features for aftershock damage potential prediction through secondary XGBoost modelling. The research results demonstrated that the amplitude, spectral characteristics, and the duration of both the mainshock and aftershock were comprehensively incorporated by the XGBoost algorithm. Consequently, a more precise assessment of the mainshock-aftershock sequence damage potential was achieved, and the model interpretability enhanced was demonstrated by the predicted values.