Graphene/SiO₂ Nanoflakes as Reliable Bio-Photonic Synapse Integrating with Dual-Functionalities of Sensing and Memory Effects for Energy-Efficient Artificial Neural Networks

In recent years, the system of artificial intelligence is still suffocated by the traditional von Neumann computing architecture. This inherent limitation has led to the bottleneck of computing speed and the problem of high energy consumption that needs to be imperatively solved. To ameliorate von Neumann bottleneck in conventional computing, we began to adopt a neuromorphic computing model, which is inspired by the neural network in human brain and allows the system to actualize “Computing in Memory”. Herin, the photonic synapse based on graphene/PMMA@silica nanoflakes is demonstrated to emulate fundamental synaptic behaviors endowed with sensing and memory functionalities. The bio-photonic synapse proposed in this work offers a series of fundamental synaptic functionalities, such as excitatory postsynaptic current (EPSC), paired-pulse facilitation (PPF), and short-term plasticity (STP) to long-term plasticity (LTP) conversion under weak light stimulation. It is found that the resulting PPF index of our graphene-based bio-photonic synapse stimulated by 365 nm light is 267.3% when the interval time is 0.1 s under the employment of -3 V in gate voltage. Additionally, after 24 pulses of light stimulation at a frequency of 0.40 Hz, the relaxation time constant still reaches up to 193.7 s, indicating the critical synaptic plasticity of our devices. The plasticity can be flexibly modulated by the gate voltage, which further contribute to a dynamic synapse with controllable weight. The learning ability of our bio-photonic synapse is confirmed by simulating the learning process with distinct moods via adjusting the gate voltages. Furthermore, due to the charge-trap rich center in silica nanosheets, the bio-photonic synapse can also act as non-volatile flash memory, well mimicking the long-term plasticity (LTP) and keeping long retention time. It should be pointed out that the implement of light-stimulated photonic synapses via the strategy of heterostructure engineering represent a potential avenue that allows the ultrafast processing speed, high bandwidth and low crosstalk than electrical-stimulated synapses. This bifunctional-photonic synapse is anticipated to bridge the gap between the neuromorphic computing and sensing data in our life and pave the way for "Computing in Memory", promoting the emerging artificial neural networks.