Recently, the prestigious international journal Nature Communications published online the latest research findings of the team led by Professor Yang Rui and Professor Miao Xiangshui from the School of Integrated Circuits at Huazhong University of Science and Technology. The study, titled "Compact Intrinsic Stochastic Photonic Spiking Neurons Based on Phase-Change Materials for Probabilistic Computing," aims to tackle the bottleneck of artificial intelligence computing power in the post-Moore era. The researchers propose a novel photonic neuron architecture utilizing antimony telluride (SbTe₉) volatile phase-change materials.

Traditional photonic neuron devices are often constrained by their large size, intricate structures, and deterministic responses. To overcome these limitations, the team designed a composite microstructure that incorporates an antimony matrix embedded with Sb₂Te₃ nanocrystals. By harnessing the inherent fluctuation mechanism of melting point and temperature during the phase-change process, the researchers achieved stochastic spiking characteristics without the need for external entropy sources.

Experimental results reveal that the device boasts an active area of merely 1.5 μm² and demonstrates a light modulation capability of 2 dB/μm. Moreover, it achieves a rapid response speed of 660 ns. In Bayesian inference tasks related to breast cancer diagnosis, the device attains an impressive accuracy rate of 98.67%. This performance significantly enhances the robustness of fuzzy sample recognition compared to traditional deterministic architectures. Additionally, in image recognition tasks, the accuracy degradation caused by synaptic programming deviations is reduced by an order of magnitude compared to deterministic devices. Under input noise conditions, the device only experiences a 4.28% loss in accuracy, showcasing its excellent anti-interference capabilities.

This breakthrough lays a bionics-inspired foundation for implementing native probabilistic computing at the hardware level. It is expected to propel the development of photonic neuromorphic computing towards high-density integration and low-power applications.