A Robust Biomimetic van der Waals Heterostructure Visual Neuromorphic Device for Multiscale In-Sensor Reservoir Computing

Visual neuromorphic devices are critical for processing exponentially increasing complex visual information due to their capability in capturing, storing, and processing optical signals from the environment. In particular, all-two-dimensional material heterostructure-based visual neuromorphic devices are recognized as one of the most promising device architectures for mitigating lattice mismatch and interfacial defects inherent in conventional heterostructures, owing to their atomic-scale interfacial compatibility. Herein, we report a robust visual neuromorphic device based on graphdiyne/MoS₂ all-two-dimensional material heterostructure. By utilizing the unique sp-sp² hybridization and abundant two-dimensionally distributed charge carrier interaction sites provided by acetylenic bonds in graphdiyne films, we achieved nearly a 10-fold enhancement in the memory window of the device. The device exhibits an ultrahigh on/off ratio of 5 × 10⁷, cycling endurance of 70 cycles, and 4 weeks air stability. Most importantly, a multiscale in-sensor reservoir computing system was developed, demonstrating over 90% recognition accuracy on facial data sets under 5%-40% Gaussian noise conditions. Its tunable relaxation time characteristic enables efficient motion trajectory recognition with an accuracy of 95.46%. A proof-of-concept demonstration confirms that the device achieves trajectory recognition within a temporal resolution range of 10²-10⁶ ms. This work presents a promising device codesign approach that advances the development of neuromorphic vision systems.