The response of tunnel blasting excavation is influenced by rock mass heterogeneity and initial stress fluctuations, showing strong random dispersion and limiting prediction accuracy. A systematic study on blasting vibration monitoring and drilling data acquisition is conducted to examine the existence and mechanism of a lower bound of prediction accuracy.Based on blasting cycle data, statistical analyses of peak particle velocity and overbreak(underbreak) are performed. Under identical design conditions, both responses exhibit stable distributions. The coefficients of variation are 19.99% and 14.81%, respectively. With construction optimization, response dispersion stabilizes and an error plateau emerges.A blasting response disturbance model is established using random field theory, and a relationship between prediction error variance and rock mass variability is derived. Theoretical analysis shows that prediction error approaches a non-zero lower bound governed by rock randomness, consistent with field data.A machine learning model based on drilling parameters is further developed to capture nonlinear relationships between multi-source inputs and blasting responses. Results show improved accuracy, while errors remain bounded by the identified lower limit.In conclusion, the lower bound mechanism of prediction accuracy is revealed through statistical analysis, random field theory, and data-driven modeling, providing a basis for parameter optimization and risk control.