# Low‐Power Perovskite‐Based Memristors Enable Fused Reservoir Computing and Neuromorphic Vision with Highly Accurate Color Perception

## TL;DR

This paper introduces low-power perovskite-based memristors that enable efficient neuromorphic vision systems with accurate color perception and multimodal recognition.

## Contribution

A novel perovskite heterostructure is developed for fused reservoir computing and neuromorphic vision with high accuracy and low power consumption.

## Key findings

- The device achieves 84% accuracy in recognizing multicolor handwritten MNIST images.
- The memristor exhibits ultra-low power consumption of 400 fJ per synaptic weight change under red light.
- The device's stability is maintained for 2 months with dual switching modes for memory and computing.

## Abstract

Integrating multicolor perception with neuromorphic vision systems, capable of emulating the procedures of image detection, storage, and local processing, represents a significant advancement in artificial visual technologies. However, challenges related to data fusion, system complexity, and stability must be addressed to fully realize the potential of this technology. In this work, a low‐dimensional/three‐dimensional (LD/3D) halide perovskite heterostructure consisting of Ag/LD perovskitoid/3D CsFAMA/ITO is fabricated, demonstrating excellent stability for 2 months combined with the co‐existence of two switching modes, namely volatile and non‐volatile. The former mode is leveraged to construct the nodes of the reservoir computing architecture, where the fusion rate of the electrical and optical signals is examined to achieve maximum recognition accuracy of multicolor handwritten MNIST images (84%). An ultra‐low power consumption of 400 fJ per synaptic weight change is also recorded during red light irradiation. By combining experiments with different top electrode materials and extensive Density Functional Theory calculations on metal atom diffusion and clustering in the materials of interest, key atomic scale processes are identified that underlie the switching behavior and lead to improved memory performance. The ability of the proposed device configuration to accurately carry out multimodal recognition tasks opens new possibilities for realizing biomimetic systems.

The development of low‐power multicolor neuromorphic computing systems could endow artificial vision machines with optical perception and processing capabilities in a way similar to biological systems, such as the human eye. Along these lines, perovskite‐based heterostructures emerge as a promising solution since their mixed ionic/electronic properties can be leveraged to build powerful in‐memory computing systems.