Researchers from Westlake University, Hangzhou Dianzi University and Beijing BOE Optoelectronics Technology have demonstrated a monolithically integrated, active-matrix perovskite neuromorphic photodetector array that combines high-resolution imaging with in-sensor visual processing and prediction capabilities.
The team developed a 256×256 pixel array (65,536 pixels) with a density of 254 pixels per inch, fabricated by integrating blade-coated formamidinium lead iodide (FAPbI3) perovskite films onto thin-film transistor (TFT) active-matrix backplanes. This architecture enables large-area operation with low crosstalk, addressing a key limitation of earlier dot-type or passive crossbar neuromorphic arrays that lack the scale required for high-resolution spatiotemporal imaging.
A central challenge in photovoltaic-type perovskite synaptic devices is balancing fast photoresponse with sufficient defect-mediated carrier trapping, which is necessary for synaptic weight modulation. To address this, the researchers introduced cesium lead bromide (CsPbBr3) into the FAPbI3 matrix. This compositional engineering plays a dual role: it passivates in-grain bulk defects while simultaneously increasing defect density at grain boundaries. The resulting defect distribution enhances carrier capture and release dynamics without sacrificing the intrinsic advantages of photovoltaic operation, such as low dark current and self-powered behavior.
As a result, the doped devices achieve a photoresponsivity of 0.39AW−1 at 640 nm and a response time of 48 ms under zero-bias operation. Importantly, the excitatory postsynaptic current (EPSC) gain increases by up to 44% compared to undoped devices, reflecting improved synaptic plasticity. The devices also exhibit stronger persistent photocurrent behavior, enabling memory-like retention of optical stimuli.
At the system level, the array supports both sensing and neuromorphic processing of visual information. It demonstrates contrast enhancement in gesture recognition tasks, improving classification accuracy from 74.66% to 91.58%. In dynamic scenarios, the device leverages spatiotemporal information transformation to enable trajectory prediction with a similarity score of up to 0.95. This predictive capability significantly enhances decision-making performance, increasing obstacle-avoidance accuracy from 84.33% to 100%.
The underlying mechanism is based on defect-modulated photovoltaic operation: photogenerated carriers are captured and released by engineered defect states, enabling controllable conductance changes that emulate biological synaptic behaviors such as learning, forgetting, and relearning. This allows the device to bridge the gap between photodetection and higher-level neuromorphic computation within a single, self-powered platform.
Overall, the work demonstrates that defect-engineered perovskite materials, combined with scalable active-matrix integration, can enable large-scale neuromorphic vision systems capable of real-time motion perception and prediction.
