The rapid advancement of artificial intelligence has propelled the development of photo-synapses for neuromorphic machine vision systems. Unlike conventional photodetectors that merely convert light into electrical signals, photo-synapses integrate sensing and pre-processing functions, mimicking the biological synapses in human visual systems. However, existing β-Ga 2 O 3 photo-synapses primarily rely on oxygen vacancy engineering to achieve synaptic performance, which not only reduces device stability but also leads to high weight update nonlinearity, severely limiting their applications in neural computing.
In a new paper published in Light: Science & Applications , a team led by Professor Shujie Jiao from Harbin Institute of Technology proposed a novel mechanism for photoelectric memristive regulation based on self-trapped holes (STHs). DFT calculations revealed that the large hole effective mass (40 m 0 ) in β-Ga₂O₃ cooperates with local lattice distortions to form STHs, localizing photogenerated holes around distorted oxygen atoms. The capture barrier is only 0.10 eV, while the escape barrier reaches 0.64 eV, enabling long-term hole localization that drastically reduces mobility and produces persistent photocurrent (PPC) effect. This mechanism inherently avoids oxygen vacancies, eliminating their bottlenecks of slow migration and high weight-update nonlinearity. As intrinsic defect states solely governed by lattice distortion rather than growth conditions, STHs do not vary the concentration of other point defects, overcoming the dual limitations of traditional oxygen-vacancy strategies in concentration controllability and process compatibility.
To validate this strategy, the researchers fabricated two-terminal β-Ga 2 O 3 photo-synapses. Under 252 nm ultraviolet illumination, the devices exhibited excellent short-term plasticity (STP) and transitioned to long-term plasticity (LTP) by tuning pulse characteristics. After 50 cycles of optical potentiation and electrical depression, the photo-synapses achieved a remarkably low weight-update nonlinearity of 0.42, outperforming most reported devices, with only slight sensing-time increase. These results demonstrate that STHs-based β-Ga 2 O 3 photo-synapses are free from slow ionized-defect kinetics and deliver superior sensing-computing performance.
Furthermore, the β-Ga 2 O 3 photo-synapses were integrated into neuromorphic machine vision systems, enabling diversified in-sensor computing tasks ranging from low-level image classification to high-level motion recognition. The system achieved recognition accuracies of 99.48% and 92.70% on the MNIST and Fashion-MNIST datasets, respectively. In object tracking tasks, the system maintained 100% target tracking accuracy under 60% Gaussian noise interference. Additionally, through reservoir computing architecture, the system reached a recognition accuracy of 94.94% for 10 distinct motions in the UTD-MHAD dataset.
“These results demonstrate that β-Ga 2 O 3 photo-synapses based on self-trapped holes engineering exhibit robust performance and hold significant potential for advancing neuromorphic machine vision in the era of artificial intelligence,” the scientists conclude. “The presented strategy opens new avenues for developing high-stability, low-nonlinearity photo-synapses for intelligent vision applications.”
Light: Science & Applications
Two-terminal β-Ga₂O₃ photo-synapse for diversified in-sensor computing via self-trapped holes engineering