This study aims to establish an intelligent Cyber-Physical Production System (CPPS) with closed-loop control and self-optimization of process parameters. Empowered by a deep learning (DL) agent, the digital twin (DT) simulation can focus on the most significant process routes and eliminate 92.53% of unnecessary exhaustive iterations. On the other hand, DT can provide the most accurate simulation result for fast convergence to the optimal parameters of DL and no need to waiting for the result after the process completed. Though collaborative and autonomous interactions between DT and DL have been proposed recently, the practical integration of DT-DL and CPPS in the Electronic Manufacturing Service (EMS) industry has yet to be validated in official publications. With such novelty approach, the improvement of the production throughput (T/P) and rolled throughput yield (RTY) is 7.513% and 0.86%, respectively, which is on top of the contributions from traditional industrial engineering (IE) productivity tools.