As an emerging information technique, digital twin has made great breakthroughs in low-cost and high-efficiency physical equipment dynamic modelling and controlling. To comply with time-varying optical path states and service requirements in dynamic networks, a digital twin architecture is proposed to enable the dynamic programmable optical transceiver (POT) configuration through deep reinforcement learning (DRL). The proposed POT configuration approach is capable of establishing dynamic POT controlling model and configuring multiple optical transceivers intelligently according to the real-time network performance feedback. Different from typical POTs, digital twin has special advantages in continuously updating configuration model by learning from the real-time monitored data. Simulation results show that the proposed POT can reduce 7.0% blocking probability and improve 15.5% throughput compared with classic POTs based on static controlling criteria and neural networks.