Strain‐Engineered Zener Tunneling Diodes Enable Energy‐Efficient Neuromorphic Edge Vision

Abstract

Machine vision systems face severe bandwidth‐latency bottlenecks and high power consumption due to massive unstructured data transfer. Edge computing via sensor‐level nonlinear computation offers a promising pathway to reduce data volume. However, existing breakdown‐based neuromorphic photodiodes typically rely on high‐bias avalanche processes, resulting in substantial operating power consumption. Here, we propose a strain‐engineered Zener tunneling diode (S‐ZCD) based on a stretched graphene/silicon heterojunction, where tensile strain opens an effective graphene bandgap and enables controllable cross‐bandgap quantum tunneling. The S‐ZCD realizes a light‐modulated breakdown process along the carrier transport pathway, exhibiting ultrasharp nonlinear switching characteristics with a subthreshold swing of 10–35 mV dec −1 , a minimum breakdown threshold voltage of 0.59 V and ultra‐low operating power consumption of ∼0.1 nW at room‐temperature. Furthermore, the coupled optical‐electrical regulation enables adaptive nonlinear preprocessing for edge vision applications, including fixed‐step filtering and bias‐programmable feature extraction. The S‐ZCD suppresses redundant background information, maintains >90% recognition accuracy while reducing inference energy consumption by ∼43%, and enables device‐characteristic‐guided array‐level contour‐extraction validation.