Photon-driven synapse advances low-power neuromorphic systems

Modern artificial intelligence systems rely on moving large amounts of data between memory and processors, a design that limits speed and increases energy use. The human brain works differently: it combines memory and computation within synapses, allowing fast, efficient learning and perception. Replicating this approach in hardware is a central goal of neuromorphic computing, especially for tasks like vision, where most real-world information is gathered and processed. 

In that context, researchers have developed a new type of artificial synapse that operates entirely with light. Unlike most existing devices, which still depend on electrical signals at some stage, this system uses optical signals both to receive information and to update its internal state. Removing electrical conversion steps could lower energy use, reduce noise, and enable faster processing, particularly in vision systems that already rely on light.

As reported in Advanced Photonics, the device is built from a rare-earth-doped crystal that emits a persistent afterglow after being illuminated. This material can store optical information in the form of trapped charge carriers. When light excites the crystal, some of these carriers emit light immediately, while others remain trapped and are released later. The balance between these pathways depends on the history of illumination, allowing the material to mimic how biological synapses change strength based on past activity.

To explain and predict this behavior, the researchers developed a model that tracks how excited carriers are generated, trapped, and released over time. The model shows how prior light exposure changes the availability of trapping sites, shaping the device’s response to subsequent signals. This history-dependent behavior is central to short-term synaptic plasticity in the brain and is reproduced here without any electrical control.

Experiments confirmed two key synaptic functions. Under ultraviolet light, the device shows paired-pulse facilitation: a second light pulse produces a stronger output if it follows closely after the first. This occurs because earlier excitation partially fills trap states, increasing the fraction of carriers that emit light immediately. Under near-infrared light, the opposite effect appears. The first pulse empties trapped carriers, so a second pulse produces a weaker response, known as paired-pulse depression. Together, these effects allow the device to both enhance and suppress signals, a requirement for realistic neural behavior.  

The experimental measurements matched the predictions of the model. The researchers also showed that the device response can be tuned by adjusting light intensity, pulse duration, and timing. Additional tests confirmed that the observed behavior arises from the material’s trap states rather than simple lingering emission, strengthening the physical basis of the approach.

To explore practical use, the team combined the material with a standard silicon imaging sensor to create a prototype neuromorphic camera. In this system, the light-sensitive layer processes images as they are captured. Strong optical signals persist longer than weak ones, while noise fades more quickly. As a result, the device can enhance contrast and reduce noise directly at the sensor, without separate processing steps.

This in-sensor processing improved performance in image recognition tasks. When the researchers modeled a neural network using the measured behavior of the optical synapse, it achieved 95.99% accuracy on handwritten digit classification after denoising. Without this built-in noise suppression, accuracy dropped to about 78%. The result shows that integrating sensing and processing can improve outcomes compared with conventional workflows that treat them separately.

The current device operates on timescales of milliseconds to seconds, slower than electronic components but similar to the timing of biological visual processing. The authors suggest that both speed and energy use could be improved by scaling down the device and modifying the material properties.

The study demonstrates a path toward fully optical computing elements that combine sensing, memory, and processing in a single platform. Such systems could be useful in applications that require efficient handling of visual information, including imaging, robotics, and edge devices where power and speed are constrained.

For details, see the original Gold Open Access article by Y. Yan et al., “Fully photon-modulated synaptic devices with bidirectional plasticity for neuromorphic vision and recognition," Adv. Photon. 8(4) 046001 (2026), doi: 10.1117/1.AP.8.4.046001