With the advent of the big data era, traditional artificial intelligence hardware systems based on complementary metal-oxide-semiconductor (CMOS) technology and the von Neumann architecture are severely constrained by the physical limits of Moore's Law when facing exponentially growing computational demands, encountering significant challenges such as bottlenecks in computational efficiency improvement and high energy consumption. Consequently, integrated photonic neural networks (PNNs), leveraging the inherent advantages of photons—ultra-high speed, wide bandwidth, high parallelism, and low power consumption—have emerged as a highly promising new computing paradigm that integrates neuromorphic and photonic computing. In biology, neurons do not perform deterministic calculations; internal noise, such as the stochastic opening of their ion channels, endows biological neurons with inherent stochastic dynamics, thereby providing crucial support for probabilistic computations like Bayesian inference. Therefore, introducing probabilistic models into photonic hardware is of great significance for suppressing the overconfidence of deterministic reasoning and enhancing the accuracy and robustness of neural networks.

However, existing photonic neurons generally suffer from either large physical dimensions or weak optical responses, severely limiting high-density integration and practical application of the system. More critically, most reported photonic neurons exhibit deterministic dynamics and lack inherent stochasticity. To achieve probabilistic computation, they often rely on complex external entropy sources or feedback optical paths, which greatly increases the architectural complexity and operational overhead of photonic neural networks. Although traditional chalcogenide phase-change materials have shown great potential for constructing non-volatile multi-state photonic synapses, their non-volatile nature makes it difficult for devices to spontaneously return to the resting state after triggering pulses when used as neurons. These non-ideal factors and architectural defects collectively constitute the core challenges currently restricting the development of large-scale, low-complexity, high-performance chip-scale photonic probabilistic neuromorphic computing systems.

To address the above issues, a team including Huazhong University of Science and Technology conducted systematic research using ZEPTOOLS' in-situ TEM measurement system. The team developed the first compact chip-scale photonic spiking neuron utilizing a novel SbTe₉ phase-change material. By leveraging the stochasticity inherent in the material's microstructural evolution, this neuron achieves stable probabilistic spiking, providing a low-complexity and highly robust hardware solution for large-scale photonic probabilistic computing systems.

Title：Compact photonic spiking neuron with inherent stochasticity based on phase-change material for probabilistic computing
Journal：Nature Communications
URL：https://doi.org/10.1038/s41467-026-73070-8

The research team developed a compact chip-scale photonic spiking neuron based on a novel phase-change material, SbTe₉, leveraging the microstructural evolution during non-equilibrium melting and crystallization to achieve inherent stochasticity. Unlike traditional phase-change materials, SbTe₉ is crystalline at room temperature. When heated above its melting point by external stimulation, it transforms into a liquid state; upon removal of the stimulus, it spontaneously recrystallizes. This process of rising and falling mimics the membrane potential dynamics of biological neurons. Due to the stochastic reconstruction of the microstructure during cyclic processes, the material's melting point and local temperature exhibit inherent fluctuations. This physical-layer stochasticity enables the neuron to exhibit stable probabilistic spiking characteristics. This mechanism eliminates the reliance on external entropy sources, laying the physical foundation for constructing high-integration photonic probabilistic computing systems.

To accurately analyze the physical origin of this stochastic characteristic, the researchers systematically characterized the phase-change behavior of the SbTe₉ thin film using advanced measurement techniques. Utilizing the in-situ sample holder provided by ZEPTOOLS, the team was able to observe the real-time dynamic microstructural evolution of the SbTe₉ thin film at different temperatures. The experiments directly observed the distinctly different transformation behaviors of the Te phase and the Sb₂Te₃ phase, particularly the localized melting of the Te phase, grain boundary migration during heating cycles, and the random formation of vacancies during crystallization. These direct pieces of evidence from in-situ characterization not only verified the existence of spatial stochasticity within the material but also provided accurate scientific support for understanding how the photonic neuron converts microscopic evolution into macroscopic probabilistic spiking.

This photonic spiking neuron, while maintaining advanced functionality, demonstrates excellent hardware performance parameters and integration potential. Its active region measures only 1.5 square micrometers, comparable to the size of a single-mode waveguide, achieving a volume reduction of approximately four orders of magnitude compared to traditional non-volatile phase-change material neurons. In performance testing, the device achieved a strong modulation capability of 2 dB per micrometer, with a response time reduced to 660 nanoseconds. Furthermore, the driving voltages of nine arrayed devices exhibit a Gaussian-like distribution with a mean of 3.04 V, demonstrating excellent process uniformity, and the device shows a cycling endurance exceeding ten thousand pulses. This compact, efficient, and uniform device structure enables the dense integration of complex probabilistic neural computing networks on large-scale silicon photonic chips.

By constructing a probabilistic spiking neural network based on this photonic neuron, the research team successfully demonstrated its application value in uncertainty quantification and highly reliable decision-making. In an unsupervised learning task using a breast cancer cell dataset, the network achieved a diagnostic accuracy of up to 98.67% and was able to accurately quantify the confidence of the classification results through predictive entropy. Compared to traditional deterministic neural networks, this system can capture the ambiguity of features through spike distribution, significantly reducing the risk of misdiagnosis when samples are ambiguous. Furthermore, in image recognition tasks, the network exhibited a strong tolerance to synaptic programming deviations and input noise, with robustness several times higher than that of deterministic neurons. This demonstrates the significant advantages of photonic computing based on physical stochasticity when processing real, complex data.

As a high-end precision instrument company based in China, ZEPTOOLS is a leading provider of in-situ electron microscopy characterization solutions. Its PicoFemto series of in-situ transmission electron microscopy characterization solutions have successively provided technical support for major research findings of users both domestically and internationally. The following image shows the ZEPTOOLS in-situ TEM sample holder product used in this research achievement: