Sub-fJ programming in ferroelectric reconfigurable homojunctions for photovoltaic in-sensor computing

Abstract Conventional vision systems encounter data bottlenecks and high-power consumption in emerging applications due to the separation of sensing and computation. In-sensor computing architectures address this limitation by integrating reconfigurable, self-powered photodiodes at the pixel level to enable in-situ visual preprocessing. However, existing devices are constrained by high programming energy, poor weight retention, or complementary metal-oxide-semiconductor (CMOS) incompatibility, hindering simultaneous optimization of power efficiency, speed, stability, and integrability. Here, we demonstrate a self-powered reconfigurable photodiode based on a bipolar WSe2 channel and a sub-20-nm ferroelectric HfxZr1−xO2 (HZO) layer. The device employs a split-gate architecture to generate polarity-switchable short-circuit photocurrent under photovoltaic mode, achieving ultralow programming energy (< 1 fJ), a switching speed of 50 μs, and weight retention exceeding 100 s. When deployed as a physical convolution kernel, the device performs in-sensor matrix-vector multiplication on incident light. In simulated edge-detection tasks, it achieves a remarkably low normalized mean squared error (~ 3.2 × 10−4), producing edge maps nearly indistinguishable from ideal software results. This work establishes an energy-efficient and self-driven hardware platform that unifies sensing, memory, and computation, realizing a practical framework for in-sensor computing.