To understand why this matters, it helps to consider how conventional machine vision actually works. A typical pipeline consists of separate, serially connected stages: a lens forms an image, an image sensor digitizes it, the raw data are transmitted over a bus or network, and an electronic processor finally extracts meaning from the numbers. Every stage in that chain carries the full weight of the previous one. A high-resolution camera may generate megabytes of data per frame, yet the downstream task, whether recognizing a pedestrian or grasping an object, may need only a tiny fraction of that information. Digitizing and moving all those raw pixels consumes energy and time at every step, and the burden grows precisely where it is least affordable: on robots and autonomous platforms that must perceive a changing environment while operating under tight power, thermal and space constraints.

Optical pre-sensor computing attacks this bottleneck at its origin. The central insight is that light traveling through a well-designed optical system is already performing mathematics for free. When wavefronts diffract, interfere, scatter or pass through engineered elements, the resulting intensity pattern at the sensor plane can correspond to a convolution, a correlation, a Fourier transform or a derivative of the input scene. In other words, the optical front end can extract task-relevant features or compress visual information before the sensor ever captures an image. The sensor then records not the raw scene but a pre-processed, information-rich representation, dramatically reducing redundant data acquisition and the subsequent electronic processing load. The review surveys these elementary optical operations in detail, including convolution and differentiation, and traces how they can be composed into more powerful computational primitives.

Building on these primitives, researchers have constructed optical neural networks, architectures that map the layers of an artificial neural network onto physical light propagation. The review categorizes several distinct design families. Diffractive optical neural networks stack patterned layers that sculpt the phase of passing wavefronts, so that a network’s weights are physically encoded in the layer geometry. Fourier-optics schemes exploit the natural ability of a lens to perform the optical Fourier transform, enabling convolutions to be executed as simple pointwise multiplications in the frequency domain. Other architectures draw on scattering media and geometrical optics, and some designs are directly integrated with image sensors, blurring the boundary between optics and detector. Each family trades off fabrication complexity, alignment tolerance and computational expressiveness, and the review systematically maps this design space for readers evaluating the technology’s maturity.

One of the review’s central technical themes is nonlinearity. Linear optical systems, however elaborate, can only perform linear transformations on the input light, and deep neural networks derive much of their power from nonlinear activation functions inserted between layers. Extending optical neural networks beyond linear operations therefore requires nonlinear activation mechanisms, and the authors devote careful attention to this challenge. They compare optoelectronic approaches, in which light is converted, modulated and re-emitted to introduce nonlinearity, with all-optical mechanisms that seek nonlinearity within the optical domain itself. The comparison is framed around practical trade-offs in response speed, illumination requirements and power consumption, the very parameters that determine whether a given nonlinear scheme can support real-world visual processing rather than a laboratory demonstration.

The review then turns to the requirements that matter for deployment in physical AI, the class of embodied systems, from industrial robots to autonomous platforms, that must sense and act in the physical world. Three requirements stand out. First, compact integration: bulky free-space optical benches are incompatible with a robot’s head or a vehicle’s sensor suite, so optical front ends must be miniaturized, potentially using metasurfaces and photonic integration. Second, compatibility with real-world incoherent light: many laboratory demonstrations rely on coherent laser illumination, but natural scenes are lit by incoherent, broadband light, and practical pre-sensor systems must work under such conditions. Third, low-power reconfigurability: a fixed optical function is of limited use when tasks change, so the computational function of the optical front end should be tunable to match the target task. Active modulators provide this flexibility, but the authors caution that their power demands must be counted honestly when assessing overall system efficiency, since an energy-hungry modulator could erase the savings the optical front end delivers.

Robustness adds a further layer of engineering difficulty. Optical front ends mounted on moving platforms must withstand motion, vibration and environmental change without losing their calibrated computational function, and their operations should be matched to the specific perceptual tasks the system must perform. A pre-sensor design optimized for one visual task may be poorly suited to another, which makes co-design between the optical front end and the downstream task a central theme. The review’s treatment of these constraints is notably candid: rather than presenting optical pre-sensor computing as a finished technology, it frames the field as a set of promising building blocks whose practical value depends on solving integration, stability and power problems that are well defined but far from trivial.

Perhaps the most forward-looking portion of the review concerns how pre-sensor computing could connect to the rest of the computational stack. The authors discuss combining optical processing before the sensor with computation within and near the sensor itself, and with integrated photonic computing chips that can carry out deeper inference on the optical front end’s output. In such a coordinated perception pipeline, light would be progressively transformed and compressed at every stage: features extracted in free space or on-chip optics, further processed in sensor-adjacent circuitry, and finally interpreted by photonic or electronic neural processors. This vision of an end-to-end optical-electronic inference chain is compelling, but the authors are clear about the obstacles. Coupling losses between optical stages, the need for compatible signal representations across the optical-electronic boundary, and the challenge of jointly training heterogeneous optical and electronic components all remain open problems that the community must solve before such pipelines become routine.

Taken together, the review positions optical pre-sensor computing as a potential visual foundation for physical AI, a way of delivering the perceptual information that subsequent planning and action modules need, at a fraction of the data movement and energy cost of conventional pipelines. The significance extends beyond any single application. As machines are asked to perceive faster, richer and more power-constrained environments, from autonomous vehicles navigating cluttered streets to drones inspecting infrastructure, the traditional separation between imaging and computing becomes an increasingly expensive artifact of history. The Tsinghua team’s synthesis suggests that the boundary is now negotiable, and that some of the earliest, cheapest and most parallel computation available to a machine may be the computation that light performs on its own way to the sensor. For a field racing to give physical AI systems fast, frugal and reliable eyes, that is a proposition worth watching closely.

Subject of Research: Optical pre-sensor computing for efficient machine vision in physical AI

Article Title: Optical pre-sensor computing opens new possibilities for efficient vision in physical AI

Article References: Optical pre-sensor computing opens new possibilities for efficient vision in physical AI. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: optical pre-sensor computing, optical neural networks, machine vision, physical AI, diffractive optics, Fourier optics, nonlinear activation, image sensors, photonic computing, low-power perception, robotics, Tsinghua University

In a review published in Frontiers in Photonics, Ranjeet Dwivedi, Charles Pourny, Anthony Albanese and colleagues from INSA Lyon, CNRS-Université Bourgogne Europe, CEA-LETI and the University Grenoble Alpes survey recent advances in photonic devices that integrate phase-change materials, with an eye toward neuromorphic computing—computing architectures inspired by the brain. Their central claim is bold: PCMs are not merely an optional tuning layer on photonic chips, but a key enabling material that bridges photonics’ high-throughput compute capability with the dense, persistent, low-energy programmability that practical AI hardware demands.

Photonics has long been touted as an attractive route around the energy problem. Optical signals propagate with very high bandwidth, low latency, and inherent parallelism through wavelength, spatial, and temporal multiplexing. These properties have motivated optical neural networks, photonic matrix-vector multiplication engines, and photonic tensor cores as potential accelerators for the linear-algebra operations that dominate inference and signal-processing workloads. But a central challenge remains: practical photonic processors require dense, programmable, reliable, and energy-efficient memory elements to store weights, biases, routing states, and calibration parameters directly inside the photonic circuit. Thermo-optic or free-carrier-based photonic circuits must be continuously powered to hold a given configuration, creating a static energy overhead that grows with system size.

Phase-change materials solve this missing memory problem in a striking way. Chalcogenide PCMs such as GeSbTe (GST), GeSbSeTe (GSST), Sb2S3 and Sb2Se3 exhibit large, reversible changes in optical refractive index and absorption between their amorphous and crystalline states. Crucially, these states are nonvolatile: once programmed, the optical state is retained without continuous electrical power. This makes PCM-based devices especially relevant for photonic computing, where many synaptic weights must remain fixed during inference while optical signals perform matrix-vector multiplication operations at high speed. Demonstrations have already shown PCM-based photonic in-memory computing, all-optical neurosynaptic systems, integrated photonic memories, and photonic tensor cores combining PCM weight banks with wavelength multiplexing.

The review begins with the materials science. Metavalent bonding (MVB) has been identified as a distinctive bonding mechanism that characterizes the crystalline phase of PCMs, placing them between covalently bonded and metallically bonded materials. Materials exhibiting metavalent bonding possess a characteristic set of features, including a high optical dielectric constant, enhanced chemical bond polarizability reflected by large Born effective charges, strong lattice anharmonicity associated with high Grüneisen parameters, and moderate electrical conductivity. From an optical perspective, PCMs exhibit significant differences between their amorphous (covalent) and crystalline (metavalent) phases, which is why they have long served as the working medium of optical data storage technologies such as CD-RW, DVD-RAM, and Blu-ray disks. The amorphous phase is obtained by heating the material above its melting temperature and then rapidly quenching it; the crystalline phase is achieved by annealing above the crystallization temperature, followed by slow cooling.

Among all PCMs, Ge2Sb2Te5 (GST225) is the most extensively studied. Located on the GeTe–Sb2Te3 pseudo-binary line of the Ge–Sb–Te ternary phase diagram, GST225 crystallizes at approximately 150–180 °C and offers a large refractive index contrast across the visible, near-infrared, and mid-infrared spectral ranges. At the telecommunication wavelength of 1550 nm, however, crystallization also increases the extinction coefficient by about an order of magnitude, leading to high optical losses. To evaluate and compare different PCMs for photonic applications, the researchers employ a material figure of merit (FOM) that quantifies the trade-off between optical modulation efficiency and insertion losses. Conventional GST225 thin films exhibit a FOM of approximately 2 at 1550 nm, while GeTe reaches 6.2 and Sb2S3 4.2. A distinct class of ultra-low-loss PCMs such as Sb2S3 and Sb2Se3 keeps the extinction coefficient close to zero in both states, though with a substantially smaller refractive index contrast attributed to their predominantly covalent bonding character.

The most intriguing materials lie at the border between covalent and metavalent systems: the GeSe–GeTe pseudo-binary line (GeSe1-xTex alloys). This alloy series bridges GeSe, characterized by a covalently bonded orthorhombic crystalline structure and ultra-low optical losses, and GeTe, a prototypical phase-change material whose rhombohedral crystalline phase is governed by metavalent bonding. Most GeSe1-xTex compositions exhibit a large refractive index contrast with limited absorption changes, resulting in FOM values up to 550 percent higher than GST225 and up to three times greater than GSST2241. Nitrogen-doped GeSe1-xTex thin films perform even better: N-doped GeSe0.5Te0.5 combines a FOM of 11.6 at 1550 nm, a large refractive index contrast, low optical losses, and a high crystallization temperature of about 345 °C. The material properties can be tailored by adjusting the Se/Te ratio and nitrogen content during deposition. As a bonus, all these alloys exhibit exceptionally high nonlinear Kerr coefficients and negligible two-photon absorption, outperforming many conventional nonlinear photonic materials and making them attractive for supercontinuum and frequency-comb generation.

Turning to devices, the review presents novel photonic designs for electrical PCM actuation. A key requirement for advanced photonic computing is the ability to encode multiple discrete phase levels, typically achieved through partial amorphization of the PCM layer, where intermediate refractive indices correspond to intermediate phase shifts. Conventional designs based on uniform heaters suffer from abrupt phase transitions because small variations in pulse parameters can produce disproportionately large phase shifts. To overcome this, the researchers propose a segmented heater architecture in which the heater width varies along the propagation direction, introducing a controlled thermal gradient so that different segments of the PCM experience distinct degrees of amorphization. When averaged across multiple segments, the cumulative effect is a smooth overall amorphization and a nearly continuous phase-shift response. This approach enables up to 164 distinguishable phase levels—greater than 7-bit resolution—with low insertion loss of approximately 0.6 dB in the worst case, while complying with standard foundry constraints (minimum heater width of 1 µm with incremental increases of 50 nm). The average power dissipation across the segmented heater is approximately 0.94 W, with programming pulse durations from 142 to 1015 ns, corresponding to an energy consumption ranging from 133 nJ to 955 nJ.

The researchers also present a hybrid-mode interferometer (HMI) that directly integrates phase-change modulation with modal interference in a single compact structure. The HMI exploits interference between hybridized transverse modes in a silicon waveguide integrated with a PCM layer at 1550 nm. By utilizing the vertical asymmetry of the structure, the waveguide is engineered to support hybridization between the fundamental TM mode and the first higher-order TE mode. In the crystalline state, the PCM-integrated waveguide length is selected to allow constructive interference, providing high transmittance with ultra-low insertion loss. In the amorphous state, nearly equal modal amplitudes arise from near-perfect hybridization, resulting in very low transmittance because the phase difference approaches destructive interference. This mechanism enables a high extinction ratio of about 14 dB within a compact footprint of roughly 24 µm length, with transmittance varying from 0.98 to 0.04—suitable for analog encoding and multi-bit storage with sub-microsecond switching speeds. Compared with conventional architectures, the HMI provides a good trade-off between compactness, multilevel encoding capability, and low insertion losses: MZI-based PCM phase shifters have large footprints (hundreds to thousands of square micrometers) because of input/output couplers and cross-talk constraints, while ring resonators, though compact, encode only a limited number of levels—for example, up to only 14 distinct optical levels demonstrated with a graphene-actuated Sb2Se3-based ring resonator.

Finally, the review presents experimental results for optically actuated GST225-based waveguides on silicon-on-insulator. GST225 patches integrated onto SOI waveguides were characterized under all-optical programming and readout conditions, with propagation losses of approximately 5 dB/µm at 1550 nm. A heterodyne interferometric setup—using a short 400 ps amorphization pulse followed 25 µs later by a longer, less intense recrystallization pulse, with a low-power 0.6 mW probe laser—monitored transmission and phase variations in real time. For a 1.5 µm-long GST225 patch, 400 ps amorphization pulses with an energy of approximately 100 pJ produced a transmission increase of about 60 percent and a phase-delay reduction of 55°. The write–erase sequence, using optimized pulse energies of 100 pJ and 7.4 nJ, was repeated over a 14-hour cycling test, providing a quantitative assessment of reversibility and endurance under repeated optical programming. The researchers conclude that the combination of devices based on electrical and optical actuation using different PCM compounds can provide a path toward achieving both static synaptic weighting (electrically) and dynamic synaptic reconfigurability, signal summation, and activation functions (optically)—the missing ingredient to enable next-generation computing systems where all the key functions can be addressed by either optical or electrical stimuli, depending on the exact required function.

Subject of Research: Physics

Article Title: Photonic devices based on phase-change materials for neuromorphic computing

Article References: Dwivedi, R., Pourny, C., Albanese, A., Yadav, A., Orobtchouk, R., Benane, G., Sawant, R., Tomelleri, M., Brulé, Y., Rogemont, A., Coillet, A., Charbonnier, B., Noé, P., Cluzel, B., & Pavanello, F. (2026). Photonic devices based on phase-change materials for neuromorphic computing. Frontiers in Photonics, 7, Article 1918762. https://doi.org/10.3389/fphot.2026.1918762

Image Credits: AI Generated

DOI: 10.3389/fphot.2026.1918762

Keywords: phase-change materials, photonics, neuromorphic computing, photonic computing, chalcogenides, optical memory, silicon photonics, Photonic, devices, based, phase-change, materials

Diffractive neural networks, sometimes called diffractive deep neural networks, are built from a sequence of thin layers whose transmission or reflection coefficients are optimized by a computer. When a light wave passes through these layers, each point on one layer diffracts light toward many points on the next layer, and the pattern of connections between points behaves like the weights of an artificial neural network. Because the connection strength is set by how light spreads and interferes, the entire forward pass of the network happens at the speed of light. In prior demonstrations, however, the input was typically a static two-dimensional image or a single spatial pattern, and the layers were designed to transform that pattern from one spatial plane to the next. Such systems excel at tasks like image classification, but they treat every input as frozen in time.

The new approach recognizes that real-world signals, from communications waveforms to biological dynamics, are inherently temporal, and that a beam of light carries information in multiple degrees of freedom simultaneously: its spatial distribution, its wavelength, its polarization, and its temporal profile. The researchers show that by manipulating the spatiotemporal light field, meaning the way the field’s amplitude and phase evolve across space and time together, a diffractive network can encode, transform, and classify information that changes over time, all within the optical domain. The network layers are no longer merely spatial masks; they become spatiotemporal operators that shape how different temporal components of the input interfere and propagate.

Technically, the framework treats the optical field as a function of both position and time, and the diffractive layers are optimized so that the light field emerging from the final layer encodes the desired output, for example a classification decision or a transformed waveform. The design process draws on numerical modeling of wave propagation combined with training procedures familiar from machine learning, in which the layer parameters are iteratively adjusted to minimize an error function. Once training converges, the learned parameters are physically implemented in the optical layers, and the network performs inference passively, with no computation performed electronically during operation. This all-optical inference path is what distinguishes the architecture from hybrid optical-electronic schemes, where light performs some operations but electronic processors handle the rest.

The significance of adding the temporal dimension is substantial. A purely spatial diffractive network processes each snapshot independently, so it cannot natively recognize patterns that unfold over time, such as a spoken word, a sequence of pulses in a fiber, or the changing intensity of a dynamic scene. A spatiotemporal diffractive network, by contrast, can in principle integrate information across a temporal window as the light propagates, allowing the physics of diffraction and interference to perform temporal filtering, correlation, and sequence recognition. The authors present this capability as a route toward optical systems that can handle streaming data directly at the front end of a sensing or communication system, before any signal is converted to electronics.

The implications for energy efficiency are among the most compelling aspects of the research. Conventional artificial intelligence hardware consumes considerable power moving data between memory and processing units, and much of that cost is incurred performing the matrix multiplications that dominate neural network inference. Diffractive optical networks perform those multiplications passively, as light diffracts and interferes, so the energy cost of the forward pass is largely limited to the energy used to generate and detect the light. By extending the architecture to spatiotemporal operation, the new framework broadens the class of problems that can benefit from this efficiency, potentially including ultrafast signal processing in optical communications, where data streams already exist as modulated light and never need to be converted at all.

Speed is the other headline advantage. Because the computation is performed by propagating light, the latency of inference is set by the time it takes the wave to traverse the network, which can be on the order of picoseconds for compact devices. For temporal signals, this means the network can in principle keep pace with data rates that overwhelm electronic processors. The authors emphasize that the spatiotemporal manipulation of the light field is what unlocks this regime: by structuring the field in time as well as space, the network can perform operations on waveforms that would otherwise require high-speed sampling and digital signal processing chains.

The framework also connects to a broader research effort aimed at exploiting the full dimensionality of light for computing. Modern optical technologies can control wavelength, polarization, orbital angular momentum, and coherence, and each of these degrees of freedom can serve as a carrier of information or as a computational resource. Spatiotemporal light field manipulation, in which ultrafast pulses are shaped simultaneously in space and time, has matured rapidly in recent years, enabling phenomena such as space-time wave packets and light sheets with engineered group velocities. The new work harnesses this toolbox for neural computation, suggesting that the design space of optical neural networks is far larger than the spatial-only architectures explored to date.

As with any emerging technology, practical considerations will shape how quickly these systems move from laboratory demonstrations to deployed applications. Implementing spatiotemporal diffractive layers requires optical components that can impose carefully designed transformations on fast-varying fields, and the accuracy of the physical implementation relative to the trained model determines the network’s real-world performance. Alignment, fabrication tolerances, and detector bandwidth all matter. The authors frame their contribution as establishing the principles and design methodology for this new class of networks, providing a foundation on which experimental implementations across different spectral bands and platform technologies can be built.

The research arrives at a moment of intense global interest in unconventional computing substrates, driven by the growing energy and speed demands of artificial intelligence. Photonic approaches ranging from integrated silicon photonics to free-space diffractive optics promise orders-of-magnitude improvements in the energy efficiency of certain computations, and diffractive neural networks are among the simplest and most scalable of these approaches, since they can be fabricated as passive optical elements and require no active switching during inference. By showing that the same diffractive framework can be extended into the temporal domain, the study expands the reach of optical neural computation from static pattern recognition toward dynamic signal processing, a capability that could matter for applications as varied as ultrafast imaging, optical communications, lidar, and the analysis of fast biological processes. The work suggests a future in which the front end of an intelligent system is not a camera feeding a processor, but a shaped light field that has already done the thinking on its way to the detector.

Subject of Research: All-optical diffractive neural networks that process spatiotemporal light fields for temporal information processing

Article Title: Spatiotemporal all-optical diffractive neural networks empowered by spatiotemporal light field manipulation

Article References: Feng, F., Zhang, Z., Huo, D., Li, X., Lin, Q., Zhao, X., Hou, G., Dai, D., Somekh, M. G., & Yuan, X. (2026). Spatiotemporal all-optical diffractive neural networks empowered by spatiotemporal light field manipulation. Light: Science & Applications, 15(1), Article 382. https://doi.org/10.1038/s41377-026-02366-7

Image Credits: AI Generated

DOI: 10.1038/s41377-026-02366-7

Keywords: diffractive neural networks, all-optical computing, spatiotemporal light field manipulation, optical neural networks, photonic computing, wavefront shaping, ultrafast optics, machine learning hardware, energy-efficient computing, optical signal processing, Light Science and Applications, deep learning optics

Modern artificial intelligence systems are extraordinarily powerful, but much of their energy and time can be consumed moving data rather than performing calculations. In conventional computing architectures, sensory inputs and intermediate results must travel repeatedly between processors and external memory. This separation is especially costly for neural networks, which perform large numbers of mathematical operations on streams of data and often need to preserve information about what happened moments or even minutes earlier. The human brain takes a different approach. Neurons both process signals and maintain internal states, allowing computation and memory to coexist across densely interconnected networks. FLARE, short for a fully in-memory photonic computing architecture, is designed around a similar principle: sensing, processing and memory are integrated into the same computing fabric rather than treated as isolated stages.

The system combines photonic and electronic mechanisms to create what the researchers describe as reconfigurable photonic neurons. Photonic computing uses light to carry and manipulate information, an approach that can exploit the speed and parallelism of optical signals. Electronic components, meanwhile, can provide storage, control and nonlinear behavior that are difficult to implement efficiently with light alone. In FLARE, these mechanisms are coupled so that the artificial neurons can respond to incoming signals, transform them through a deep nonlinear neural network and preserve aspects of their previous activity. That memory is not a single, uniform function. The architecture supports both long-term and short-term dynamics, allowing it to retain learned or accumulated information while continuing to react quickly to new inputs.

This distinction between memory timescales is central to systems that must operate in changing environments. Short-term dynamics can help a machine interpret rapidly evolving signals, such as motion, acceleration or changing visual information. Long-term retention can preserve a more persistent internal state, allowing the system to use information gathered earlier rather than treating every new signal as an isolated event. The FLARE chip reportedly maintained long-term memory for 7.45 seconds while preserving short-term behavior in the gigahertz range. A gigahertz corresponds to a billion cycles per second, so the result represents a combination of sustained memory and extremely rapid signal dynamics. Rather than forcing a system to choose between storing information and processing it quickly, the architecture is intended to make both behaviors available within the same neural substrate.

At the physical level, the promise of photonic computing comes from the way light can propagate and interact across many channels at once. Optical signals can encode information in properties such as intensity, and integrated photonic circuits can perform transformations as light travels through carefully engineered structures. In a neural network, these transformations can represent the weighted combinations of inputs that form the basis of artificial-neuron computation. The critical challenge is that useful intelligence requires more than fast linear operations. Neural systems must also incorporate nonlinear responses, memory and adaptation. FLARE addresses those demands by combining optical processing with electronic mechanisms, creating neurons whose responses can be reconfigured and whose internal state can evolve over time. The result is not simply an optical accelerator attached to a conventional processor, but an attempt to merge the roles of memory, computation and sensing.

Scale is another important feature of the reported demonstration. The researchers built a monolithic multicore chip with 7,378 neurons, then arranged ten cores into a multilayer FLARE system for the autonomous-navigation experiment. A monolithic design places the relevant elements within a unified chip architecture, a step intended to reduce the bottlenecks that arise when separate components must exchange data. Multiple cores and layers allow the network to expand beyond a small laboratory proof of concept and to process information through a deeper neural structure. The architecture’s reconfigurability is also significant: a system designed to operate in the physical world must adjust to different inputs and tasks rather than relying on a fixed, one-purpose circuit. By integrating memory directly into the neural elements, FLARE is designed to keep intermediate activations close to where they are generated, limiting the costly movement of data through external digital memory.

The researchers tested the architecture in a setting that demands continuous interaction between perception and action: autonomous racing-drone navigation. A racing drone must interpret sensory information, explore its environment and adapt its behavior while moving rapidly through space. Conventional systems often divide these tasks among sensors, processors, memory and control units, creating delays and energy costs as information is repeatedly converted, transferred and processed. In the FLARE demonstration, the multilayer photonic system handled sensing, exploration and adaptation as parts of an integrated neural process. Its reported energy cost of 61.87 attojoules per operation illustrates the potential efficiency of performing computation in place. An attojoule is 10^-18 joules, an extraordinarily small unit of energy. The figure is reported at the system level, making it relevant to the complete demonstrated architecture rather than only to an isolated optical operation.

The most striking implication is that FLARE could help shift artificial intelligence away from the traditional sequence of “sense, store, compute and act.” In embodied systems such as drones, robots and autonomous vehicles, intelligence is inseparable from the physical world. Sensors generate continuous streams of data, and useful decisions depend on temporal context: what the machine detected earlier, how the environment is changing and whether a previous action produced the expected result. A network with integrated memory can preserve this context as part of its ongoing activity. A network that also processes signals photonicly could, in principle, handle high data rates without sending every intermediate result to a distant electronic memory. The combination may be particularly valuable where size, energy consumption and response time are tightly constrained.

The work nevertheless represents a pathway rather than a finished replacement for conventional artificial intelligence hardware. The reported results establish the architecture’s neuron count, memory retention, high-speed dynamics and autonomous-navigation demonstration, but practical deployment will depend on how such systems perform across a broader range of tasks and operating conditions. Photonic and electronic components must remain precisely coordinated, and large integrated neural systems must be manufactured, programmed and calibrated reliably. Memory that lasts several seconds may be useful for navigation, but different applications could require other timescales. Likewise, energy per operation is only one part of a system’s total cost; sensing, communication, control and training procedures also matter. These are engineering questions that will shape whether fully in-memory photonic systems can move from specialized demonstrations into everyday machines.

FLARE’s significance lies in the way it brings several long-standing ambitions together on one platform. It uses light for fast, parallel signal processing, electronics for memory and control, neural-network organization for nonlinear computation, and integrated architecture for direct interaction with sensory inputs. The resulting chip does not merely imitate the brain’s appearance; it targets a functional property that makes biological intelligence efficient: memory and computation are deeply intertwined. With thousands of photonic neurons, seconds-long retention, gigahertz-scale short-term dynamics and a low reported energy cost in drone navigation, the system offers a glimpse of machines that compute less by moving data and more by transforming information where it already exists. If the approach continues to scale, future autonomous devices could become faster, more adaptive and substantially more energy-efficient without relying on the memory bottlenecks that define much of today’s AI hardware.

In a new paper published in Light: Science & Applications, a team of scientists, led by Professor Yidong Huang from the Department of Electronic Engineering in Tsinghua University and their collaborators from Peking University, Berxel Photonics Company Ltd. and Shenzhen Technology University have proposed the SUANPAN architecture for optical inner product instead of optical matrix operations. Just like the transistors in an integrated circuit, the independent basic computing unit in such scheme contains only one emitter-detector pair and could be scaled up to form a photonic computing chip. The elemental values of two vectors are encoded on the output intensity of the light-emitters and the photoresponsivity of the photodetectors (PDs) by a brand-new Bit Encoding and Analog Detecting method without requiring large-scale ADC or DAC arrays. The photocurrent of the PD would be proportional to the multiplication of the light intensity and photoresponsivity, and the final result of the inner product can be obtained by the summation of all the photocurrents. Since there is no interaction among the propagating light beams of all emitter-detector pairs and only the output currents of all PDs are connected, such scheme is scalable by increasing the number of emitter-detector pairs with no additional loss or error as well as flexibly reconfigurable and programmable for different computational tasks.

As a proof of principle, the SUANPAN architecture is implemented by utilizing an 8×8 vertical cavity surface emission laser (VCSEL) array and an 8×8 MoTe2 two-dimensional (2D) material PD array. In experiment, the calculation fidelity of random vector inner product can be as high as >98% for various bit precisions (2-bit, 4-bit and 8-bit), and >95% for various vector dimensionalities (@4-bit precision). Furthermore, such implementation has been successfully reconfigured to perform two typical AI tasks, Ising machine and artificial neural network (ANN). A randomly generated 1024-dimensional Ising problem is successfully solved, which is the highest dimensionality of optical Ising machine with heuristic algorithm. Meanwhile, a competitive classification accuracy of 88% is achieved for ANN on MNIST handwritten digit dataset. It is believed that photonic SUANPAN is capable to serve as a fundamental linear vector machine and is potential to enhance the computing power for future various AI applications.

These scientists summarize the operational principles and advantages of SUANPAN:

“It breaks through the traditional mindset of obtaining optical matrix transformations through interaction of light beams. Instead, there is no interaction among those propagating light beams of all emitter-detector pairs. Therefore, the SUANPAN can be decomposed into emitter-detector pairs as independent computing units. The scalability, reconfigurability and programmability of the SUANPAN architecture are only based on the multiplication, recombination and modulation of emitter-detector pairs without any additional cost. Compared with optical matrix transformations through interaction between light beams, the SUANPAN possesses following advantages: (1) With massive and industrial multiplication of emitter-detector pairs, the SUANPAN can theoretically be infinitely scalable. (2) The SUANPAN can be flexibly reconfigured and programmed to perform various specific computing tasks. (3) Only correcting the intensity of light beam is required, and there is no requirement to correct the phase term. (4) Even if one emitter-detector pair is broken during fabrication or operation, other emitter-detector pairs would not be affected, and only the operating dimensionality would be decreased.”

“SUANPAN provides a promising solution for optoelectronic analog-digital hybrid computing. Large-scale DAC and ADC arrays are usually required in optoelectronic computing. However, with Bit Encoding and Analog Detecting paradigm, M-bit digital electronic signal is converted to analog with in a set of M emitter-detector pairs, while each emitter-detector pair only represents 1-bit information. Thus, no DAC is required. At the same time, only one ADC is required to convert the total photocurrent into electronic digital signal. Therefore, the Bit Encoding and Analog Detecting computing paradigm greatly reduces the heavy burden introduced by ADC and DAC. Actually, it is also an important issue for the scalability of the SUANPAN architecture.”

                        Journal

                        Light Science & Applications

                        DOI

                        10.1038/s41377-025-02059-7 
                        Article Title

                        SUANPAN: scalable photonic linear vector machine
            Media Contact
                                WEI ZHAO
                Light Publishing Center, Changchun Institute of Optics, Fine Mechanics And Physics, CAS
            zhaowei@lightpublishing.cn
]]>
					
		
		
		136986