Compact photonic spiking neuron with inherent stochasticity based on phase-change material for probabilistic computing.

Integrated photonic neural networks (PNNs) are at the forefront of AI computing, leveraging light’s unique properties, such as large bandwidth, low latency, and potentially low power consumption. Nevertheless, the integrated optical components are inherently sensitive to external disturbances, thermal interference, and various device imperfections, which detrimentally affect computing accuracy and reliability. Conventional solutions use complicated control methods to stabilize optical devices and chip, which result in high hardware complexity and are impractical for large-scale PNNs. To address this, we propose a training approach to enable control-free, accurate, and energy-efficient photonic computing without adding hardware complexity. The core idea is to train the parameters of a physical neural network towards its noise-robust and energy-efficient region. Our method is validated on different integrated PNN architectures and is applicable to solve various device imperfections in thermally tuned PNNs and PNNs based on phase change materials. A notable 4-bit improvement is achieved in micro-ring resonator-based PNNs without needing complex device control or power-hungry temperature stabilization circuits. Additionally, our approach reduces the energy consumption by tenfold. This advancement represents a significant step towards the practical, energy-efficient, and noise-resilient implementation of large-scale integrated PNNs.

Photonic neural networks capable of rapid programming are indispensable to realize many functionalities. Phase change technology can provide nonvolatile programmability in photonic neural networks. Integrating the direct laser writing technique with phase change material (PCM) can potentially enable programming and in-memory computing for on-chip photonic neural networks. Sb2Se3 is a newly introduced ultralow-loss phase change material with a large refractive index contrast over the telecommunication transmission band. Compact, low-loss, rewritable, and nonvolatile on-chip phase-change metasurfaces can be created by using direct laser writing on a Sb2Se3 thin film. Here, by cascading multiple layers of on-chip phase-change metasurfaces, an ultra-compact on-chip programmable diffractive deep neural network is theoretically demonstrated at the wavelength of 1.55 μm and benchmarked on two machine learning tasks of pattern recognition and MNIST (Modified National Institute of Standards and Technology) handwritten digits classification, and accuracies comparable to the state of the art are achieved.

A programmable hardware implementation of all-optical nonlinear activation functions for different scenarios and applications in all-optical neural networks is essential. We demonstrate a programmable, low-loss all-optical activation function device based on a silicon micro-ring resonator loaded with phase change materials. Four different nonlinear activation functions of Relu, ELU, Softplus and radial basis functions are implemented for incident signal light of the same wavelength. The maximum power consumption required to switch between the four different nonlinear activation functions in calculation is only 1.748 nJ. The simulation of classification of hand-written digit images also shows that they can perform well as alternative nonlinear activation functions. The device we design can serve as nonlinear units in photonic neural networks, while its nonlinear transfer function can be flexibly programmed to optimize the performance of different neuromorphic tasks.

The demand for information processing at ultrahigh speed with large data transmission capacity is continuously rising. Necessary building blocks for on‐chip photonic integrated circuits (PICs) are reconfigurable integrated low‐loss high‐speed modulators and switches. Phase change materials (PCMs) provide unique opportunities for integration into PICs. Here, the investigation of layered gallium monosulfide (GaS) as a novel low‐loss PCM from infrared to optical frequencies is pioneered, with high index contrast (Δn≈0.5) at the optical telecommunication band. The GaS bandgap switches from 1.5 ± 0.2 eV for the amorphous state to 2.1 ± 0.1 eV for the crystalline state. It is demonstrated that the reversible GaS amorphous‐to‐crystalline phase transition can be operated thermally and by picosecond green (532 nm) laser irradiation. The design of a reconfigurable integrated optical modulator on‐chip based on Mach‐Zehnder Interferometers (MZI) with the GaS PCM cell deposited on one of the arms for application is presented at the telecommunication wavelength of λ = 1310 nm, where the standard single mode optical fiber exhibits zero chromatic dispersion, and at λ = 1550 nm, where a minimum optical loss of 0.22 dB km−1is obtained. This opens the route to applications such as reconfigurable modulators, beam steering using phase modulation, and photonic neural networks.

To address the fundamental challenges of limited parallelism and escalating energy consumption inherent in traditional digital artificial intelligence (AI) hardware—stemming from the von Neumann bottleneck—photonic integrated circuits (PICs) have emerged as a pivotal solution, offering the synergistic potential of high parallelism, ultra‐high speed, and low power consumption. Here, we present a photonic neural network (PNN) based on programmable phase‐change metasurfaces, which realizes an on‐chip reconfigurable parallel computing architecture through the monolithic integration of non‐volatile programmable power beam‐splitters, higher‐order mode couplers, and multimode cross‐waveguides. The programmable power beam‐splitter achieves a high precision of 9 bits, supports TE 0 and TE 1 input modes, and operates across a broad wavelength range of 1480–1620 nm. With matching bandwidths, the mode coupler and cross‐waveguide ensure scalability for large‐scale optical networks. Fabricated on a silicon‐on‐insulator (SOI) platform and fully compatible with complementary metal‐oxide‐semiconductor (CMOS) processes, the PNN demonstrates industrialization potential. In handwritten digit recognition tasks, it achieves a classification accuracy of 98.95%, while exhibiting intrinsic advantages in speed and power efficiency. This work validates the efficacy of non‐volatile phase‐change materials (PCMs) in PNNs and presents a scalable hardware paradigm for high‐performance computing. Our findings pave the way for next‐generation photonic intelligent processors, with promising applications in edge computing and data centers.

The increasing demand for efficient hardware in neural computation highlights the limitations of electronic-based systems in terms of speed, energy efficiency, and scalability. Wavelength-division multiplexing (WDM) photonic neural networks offer a high-bandwidth, low-latency alternative but require effective photonic activation functions. Here, we propose a power-efficient and tunable-bias all-optical nonlinear activation function using vanadium dioxide (VO2) for WDM photonic neural networks. We engineered a SiN/BTO waveguide with a VO2 patch to exploit the phase-change material’s reversible insulator-to-metal transition (IMT) for nonlinear activation. We conducted numerical simulations to optimize the waveguide geometry and VO2 parameters, minimizing propagation and coupling losses while achieving a strong nonlinear response and low-threshold activation power. Our proposed device features a sub-milliwatt threshold power, a footprint of 5 μm, and an ELU-like activation function. Moreover, the bias of our device could be thermally tuned, improving the speed and power efficiency. On the other hand, performance evaluations using the CIFAR-10 dataset confirmed the device’s potential for convolutional neural networks (CNN). Our results show that a hybrid VO2/SiN/BTO platform could play a prominent role in the path toward the development of high-performance photonic neural networks.

Photonic neural networks have been developed as a hardware platform to accelerate machine-learning inference. Digital micromirror devices (DMDs) have been playing a critical role in developing a variety of photonic neural networks for their ability to manipulate millions of optical spatial modes in a 2D plane at >10 kHz frame rate. DMDs have not only enabled high-throughput machine-learning inference but also made hardware-in-the-loop training possible with photonic neural networks. In this talk, we will review the functions of DMDs in a plethora of photonic-neural-network architectures and discuss how MEMS-based technologies can enable novel photonic neural networks in future.

Photonic neural network has been sought as an alternative solution to surpass the efficiency and speed bottlenecks of electronic neural network. Despite that the integrated Mach–Zehnder Interferometer (MZI) mesh can perform vector-matrix multiplication in photonic neural network, a programmable in-situ nonlinear activation function has not been proposed to date, suppressing further advancement of photonic neural network. Here, we demonstrate an efficient in-situ nonlinear accelerator comprising a unique solution-processed two-dimensional (2D) MoS2 Opto-Resistive RAM Switch (ORS), which exhibits tunable nonlinear resistance switching that allow us to introduce nonlinearity to the photonic neuron which overcomes the linear voltage-power relationship of typical photonic components. Our reconfigurable scheme enables implementation of a wide variety of nonlinear responses. Furthermore, we confirm its feasibility and capability for MNIST handwritten digit recognition, achieving a high accuracy of 91.6%. Our accelerator constitutes a major step towards the realization of in-situ photonic neural network and pave the way for the integration of photonic integrated circuits (PIC).

Optical bistable state exhibits its great potential applications in photonic integrated circuit and photonic neural network. However, the traditional optical bistable state will be influenced by the system disorders, which are not suitable for application. In this work, we investigate the topological bistable states in a layered structure with center inversion symmetry consisting of alternating layers of high index material TiO<sub>2</sub> and low index material SiO<sub>2</sub>. In the topological mode, the electric field is highly localized in the inversion center of the layered structure (also known as the interface) and exponentially decays into the bulk. Thus, when the nonlinear permittivity is strategically introduced into those layers, nonlinear phenomena such as the bistable state appears. The finite element numerical simulations reveal that the optimal bistable state appears when the layer period is 5 with a threshold power around 1.2 W/m. Benefiting from the topological characteristics, such a bistable state persists when random perturbations are introduced into the layer thickness and refractive index. Finally, we apply the bistable states to a photonic neural network. The bistable function shows prediction accuracy similar to the classic activation function ReLu and sigmoid in various learning tasks. These results provide a novel method for inserting highly robust optical bistable states from topological layered structure into photonic neural network.

Neuromorphic photonic integrated circuits over silicon photonic platform have recently made significant progress. Photonic neural networks with a small number of neurons have demonstrated important applications in high-bandwidth, low latency machine learning (ML) type signal processing applications. Naturally an important topic is to investigate building a large scale photonic neural networks with high flexibility and scalability to potentially support ML type applications involving high-speed processing of a high volume of data. In this paper we revisited the architecture of microring resonator (MRR) -based non-spiking and spiking photonic neurons, and photonic neural networks using broadcast-and-weight scheme. We illustrate expanded neural network topologies by cascading photonic broadcast loops, to achieve scalable neural network scalability with a fixed number of wavelengths. Furthermore, we propose the adoption of wavelength selective switch (WSS) inside the broadcasting loop for wavelength-switched photonic neural network (WS-PNN). The WS-PNN architecture will find new applications of using off-chip WSS switches to interconnect groups of photonic neurons. The interconnection of WS-PNN can achieve unprecedented scalability of photonic neural networks while supporting a versatile selection of mixture of feedforward and recurrent neural network topologies.

Integrated photonic neural networks provide a promising platform for energy-efficient, high-throughput machine learning with extensive scientific and commercial applications. Photonic neural networks efficiently transform optically encoded inputs using Mach-Zehnder interferometer mesh networks interleaved with nonlinearities. We experimentally trained a three-layer, four-port silicon photonic neural network with programmable phase shifters and optical power monitoring to solve classification tasks using "in situ backpropagation," a photonic analog of the most popular method to train conventional neural networks. We measured backpropagated gradients for phase-shifter voltages by interfering forward- and backward-propagating light and simulated in situ backpropagation for 64-port photonic neural networks trained on MNIST image recognition given errors. All experiments performed comparably to digital simulations ([Formula: see text]94% test accuracy), and energy scaling analysis indicated a route to scalable machine learning.

Reconfigurable silicon photonic devices are widely used in numerous emerging fields such as optical interconnects, photonic neural networks, quantum computing, and microwave photonics. Currently, phase change materials (PCMs) have been extensively investigated as promising candidates for building switching units due to their strong refractive index modulation. Here, nonvolatile multilevel switching of silicon photonic devices with Ge2Sb2Te5 (GST) is demonstrated with In2O3 transparent microheaters that are compatible with diverse material platforms. With GST integrated on the silicon photonic waveguides and Mach‐Zehnder interferometers (MZIs), repeatable and reversible multilevel modulation of GST is achieved by electro‐thermally induced phase transitions. Particularly, the segmented switching unit of In2O3 and GST is proposed and demonstrated to be capable of producing about one order of magnitude larger temperature gradient than that of the nonsegmented unit, resulting in up to 64 distinguishable switching levels of 6‐bit precision, and fine‐tuning of the switching voltage pulses is promising to push the precision even further, to 7‐bit, or 128 distinguishable switching levels. The capability of precise multilevel phase‐change modulation is crucial to further facilitate the development of nonvolatile reconfigurable switches and variable attenuation devices as building blocks in large‐scale programmable optoelectronic systems.

With the increasing computational demands of artificial intelligence, photonic neural networks have emerged as a promising alternative to electronic accelerators, offering high speed, low latency, and energy efficiency. This paper presents a low-loss crossing waveguide-assisted silicon-based microdisk resonator (MDR) crossbar array integrated with phase change material (PCM) Ge2Sb2Se4Te1 (GSST) for optical neural network applications. The designed multimode interference (MMI) crossing waveguide features a compact footprint of 1 µm2 with a 70-degree offset, reducing the device area by a factor of 2.03 while maintaining low insertion loss (IL) and crosstalk (CT). By integrating GSST on the microdisks, we achieve nonvolatile and programmable weight control via crystallization degree modulation. A 3 × 3 MDR array is demonstrated to perform photonic matrix-vector multiplication with a computational efficiency of 1.2 × 1012 MAC/s/mm2. The scalability of the system has been further validated through image convolution tasks and fall detection tasks. This work provides a scalable and efficient architecture for on-chip photonic computing.

The phase change material (PCM) used in a building wall can reduce building energy consumption and improve the living comfort level. However, only when the temperature of the PCM layer is between phase transition temperature ranges can it have the maximum effect to increase the thermal inertia. How to select the suitable PCM according to the local weather and building wall always troubles building engineers and affects the wide use of PCM on building energy efficiency. Aiming at present situation, this paper numerically analyzed the energy conservation of building multilayer wall integrated with a PCM layer to optimize PCM thermophysical properties and wall structure under the Chengdu typical climates of winter, summer and transition seasons. The numerical simulation used the heat transfer model of the enthalpy- porosity technique, which was verified by the experiment. The results show that the PCM layer can decrease the building annual load, increase the time lag, decrease the decrement factor and improve occupants' comfort. When the phase change occurs under the suitable temperature conditions, high latent heat is conducive to decreasing the heat flow fluctuation, increasing the time lag and improving inner surface temperature stability. It is more efficient that the PCM layer is located next to the internal surface and under this condition, the optimum solidus and liquidus temperatures are 14°C and 26°C respectively, under the Chengdu climates. And the annual inner surface heat flow can be reduced up to 13.36% for the PCM layer of 30mm and 9.64% for the PCM layer of 20mm compared with no the PCM layer.

The demand for high-capacity communication systems has grown exponentially in recent decades, constituting a technological field in constant change. Data transmission at high rates, reaching tens of Gb/s, and over distances that can reach hundreds of kilometers, still faces barriers to improvement, such as distortions in the transmitted signals. Such distortions include chromatic dispersion, which causes a broadening of the transmitted pulse. Therefore, the development of solutions for the adequate recovery of such signals distorted by the complex dynamics of the transmission channel currently constitutes an open problem since, despite the existence of well-known and efficient equalization techniques, these have limitations in terms of processing time, hardware complexity, and especially energy consumption. In this scenario, this paper discusses the emergence of photonic neural networks as a promising alternative for equalizing optical communication signals. Thus, this review focuses on the applications, challenges, and opportunities of implementing integrated photonic neural networks for the scenario of optical signal equalization. The main work carried out, ongoing investigations, and possibilities for new research directions are also addressed. From this review, it can be concluded that perceptron photonic neural networks perform slightly better in equalizing signals transmitted over greater distances than reservoir computing photonic neural networks, but with signals at lower data rates. It is important to emphasize that photonics research has been growing exponentially in recent years, so it is beyond the scope of this review to address all existing applications of integrated photonic neural networks.