Ultra-high precision predictive assembly of composite parts is vital for large-scale aircraft production. The current practice of composite fuselage shape control is low efficient, non-optimal and experience dependent. We propose a machine learning based ultra-high precision quality control technique that can improve the quality and reduce the flow time. The objective is accomplished by (i) building a digital twin platform, validated by experimental data; (ii) developing a surrogate model for predictive analysis; (iii) conducting multivariable optimization to determine the optimal control of actuators. In the case study, we show that the proposed technique can achieve satisfactory prediction performance and that the automated quality control system can significantly reduce the assembly time with improved dimensional quality. This research has obtained several best paper awards. We appreciate the support from National Science Foundation and DOD MEEP program.