Bidirectional synapse based on anti-resonance modulation of photonic microring

A synaptic device is fundamental for memory and learning functions of neural networks. In this work, we demonstrated a GST (Ge2Sb2Te5)-based microring synapse with nonvolatile reconfigurable characteristics. The device shows bidirectional weight modulation during unidirectional crystallization or amorphization of GST, simulating long-term depression (LTD) and long-term potentiation (LTP). In addition, we adopted an anti-resonance pump scheme to reduce the fluctuations in pump energy coupled into the device to less than 12%. This scheme significantly enhances the programming precision of the microring synapse, achieving 24 resolvable states over 15 cycles. This work holds promise for laying the foundation for novel, to the best of our knowledge, photonic computing architectures and provides possibilities for the implementation of vision and adaptive optical neural networks that rely on bidirectional plasticity.