A Phase-Transition-Driven All-Optical Neuron with Sub-Nanosecond Nonlinear Activation

Photonic neuromorphic computing has emerged as a potent solution for artificial intelligence (AI) to transcend the computational constraints of traditional integrated circuits. Nonlinear activation functions of neurons are central to deep neural networks, one of the most powerful tools in AI programs, as they enable the learning of highly intricate mappings. Currently, the absence of high-speed, low-threshold all-optical neurons presents a tremendous challenge in the field, leading to an overreliance on optoelectronic hybrids. This reliance necessitates frequent optoelectronic conversions for data transfer between photonic and electronic systems, leading to considerable latency and increased energy consumption. To address these challenges, we leveraged the light-induced phase transition in elemental tellurium (Te) to develop an all-optical neuron, which has achieved, for the first time, an ultrafast response at the picosecond level (∼260 ps), representing a nearly two orders of magnitude enhancement in speed compared to conventional all-optical technologies. Furthermore, through an integrated hardware-software approach, we have effectively demonstrated the superiority of our Te neurons within a three-layer deep neural network. The processing of nonlinear activation operations is accelerated by a factor of 100 compared to electronic neurons, with the added benefit of further enhancement by parallel processing. Additionally, the phase-transition-driven mechanism has granted Te neurons exceptional advantages in terms of feature size, threshold energy, and enhanced performance metrics. Indeed, the proposed all-optical neuron holds immense promise for enhancing the integration density and energy efficiency of photonic neural networks. It demonstrates great potential to surmount the performance limitations imposed by electronic systems, heralding a new era in photonic computing.