In-Material Self-Adaptive Dynamic Encoding in FeO_x Optomemory Device for Real-Time Analog Signal Processing

In-sensor vision computing system as an emerging edge computing platform shows great potential to process dynamic information. However, it still requires an electric to reset weight, largely limiting its capability in processing complex signals. Here we propose FeO_x optomemory that can automatically convert its states from negative photoconductance memory (NPM) to positive photoconductance memory (PPM). The formation of neutral oxygen vacancy (V_o) from both the photogenerated electron-based iron reduction process (Fe³⁺ to Fe²⁺) and the photoelectron trapping by charged oxygen vacancy (V_o^x+) builds the NPM effect under low light dosage illumination. The decrease in V_o by the photogenerated hole-based iron oxidization process (Fe²⁺ to Fe³⁺) and the increase in V_o^x+ by Joule-heating assisted photoelectron detrapping from V_o sites causes the automatic conversion from the NPM to the PPM when the light illumination exceeds the threshold dosage (0.72 µJ/µm²). Such self-adaptive conversion from NPM and PPM enables the FeO_x optomemory to execute fully optical computing. The NPM effect provides rich echo states for dynamic feature encoding while the PPM initializes the encoded states, thus building a self-adaptive reservoir computing (RC) system, yielding a recognition accuracy of 97.93%. This work provides an emerging in-material encoding mechanism for in-sensor edge computing system.