Antiferroelectric polarization enabling physical activation in CuBiP₂Se₆ for medical image processing.

Abstract

Antiferroelectric materials, featuring field controllable antipolar ordering and reversible polarization switching, offer a promising platform for hardware efficient neuromorphic computing. The tunable polarization dynamics and layered van der Waals structure enable the multifunctional integration of sensing, learning, and computation within a single device architecture. Here, we demonstrate an antiferroelectric polarization driven diode exhibiting an extended linear operating region, which simultaneously enables physical activation and computing-in-memory. Building on the device capability, we construct an in-sensor computing system that achieves over 95% accuracy in medical image classification. We further integrate the devices to demonstrate a hardware-based activation function, attaining accuracy and training loss comparable to an ideal activation function. To enhance adaptability, we further propose a tunable activation circuit that enables linear modulation of the reverse bias slope via gain control. Overall, this work establishes a dual-functional antiferroelectric heterojunction, highlighting its strong potential for constructing optically triggered, compact, and low-power perception-computation-integrated neuromorphic systems for medical image processing.