Opto-Electrical Decoupling of Phototransistors via Light-Induced Ferroelectric Depolarization for In-Sensor Computing

Highly sensitive sensors are critical for in-sensor computing, an ultrafast and low-power machine vision technology. However, capturing sharp images without motion blur in low-light and high-speed situations remains challenging due to weak photoresponse. Here, we present a heterostructure ferroelectric phototransistor leveraging opto-electrical decoupling for fast perception and in-sensor computing. The channel is preprogrammed to a low-resistance state via ferroelectric polarization, while light modulates the drain current through light-induced ferroelectric depolarization. This mechanism enables a record-high MoTe₂-based photoresponsivity of 3.05×10⁴ A/W by optimizing the balance between depolarization and screening fields. The sensors can perceive light pulses as short as 200 μs, achieving an operating frequency of 5 kHz and an energy consumption of 74 fJ. Utilizing a light-programmable neutral point, a 3 × 3 sensor array was developed as the optical kernel for scene-specific in-sensor computing, achieving a license plate recognition accuracy of 92.4% with significantly reduced motion blur. These results demonstrate the potential of this technology for high-speed, low-light machine vision applications.