image: A KAUST-designed optoelectronic device can adapt to light, mimicking synapses and neurons for optical neuromorphic computing.
Credit: © 2025 KAUST.
Optoelectronic devices developed at KAUST that behave as either synapses or neurons, and adapt and reconfigure their response to light, could find use in optical neuromorphic information processing and edge computing[1].
The team from KAUST has designed and fabricated metal-oxide semiconductor capacitors (MOSCaps) based on the 2D material hafnium diselenide (HfSe2) that act as smart memories. The devices feature a vertical stack structure where HfSe2 is sandwiched between layers of aluminum oxide (Al2O3) and placed on a p-type silicon substrate. A transparent indium tin oxide (ITO) layer sits on top, allowing light to enter from above.
“When hafnium diselenide nanosheets are integrated into charge-trapping memory devices through solution-based processes, they enable both optical data sensing and retention capabilities,” says graduate student Bashayr Alqahtani. This allows the device to be reconfigured to sense light or store optical data after the light source is removed, depending on the bias conditions. “Our device is based on a two-terminal capacitive memory, which shows promise for device 3D stacking, paving the way for more adaptive and energy-efficient solutions,” she explains.
Experiments show that the charge trapping and capacitance of MOSCaps change with light conditions, allowing them to serve as smart memories that learn using light. As a result, optical signals can be used to train and alter the response of the device, while electrical bias signals can be used to erase the device. In particular, the team has shown that exposure to blue light with a wavelength of 465nm can reinforce or strengthen the response to red light at 635nm, a behavior known as associated learning. In the terminology of neuromorphic computing, the MOSCap acts like an artificial synapse showing both long-term potentiation (increase in synaptic response) and long-term depression (weakening of synaptic response).
”This work investigates how artificial neurons respond and adapt to optical stimuli — specifically, changes in light intensity, duration and wavelength,” explains Nazek El-Atab, who led the team. “This research is crucial for understanding these smart memories’ capabilities and improving their adaptive learning mechanisms.”
The team used these characteristics to run a simulation that predicts that a capacitive synaptic array circuit based on the devices could recognize handwritten digits from the industry standard MNIST database with an accuracy of 96%. They also show that, in principle, the adaptive sensing capabilities of a MOSCap neuron could be used to perform exoplanet detection by correctly spotting transient changes in a star’s light intensity as an exoplanet periodically passes in front of the star with an accuracy of 90%.
“These devices demonstrate in-memory light sensing capabilities that make them ideal for edge computing applications,” comments El-Atab. “They show particular promise for artificial intelligence applications where rapid processing and storage of large data volumes is essential, especially when it comes to optical data. The range of potential applications can be wide — from autonomous vehicles to virtual reality and IoT systems.”