A fibrous neuromorphic device for multi-level nerve pathways implementing knee jerk reflex and cognitive activities

Here, we present fibrous neuromorphic devices (FNDs) that serves as multi-level nerve pathways to implement a biomimetic knee-jerk reflex and cognitive activities. By the tunable charge-carrier polarity of the fibrous electrolyte, FNDs successfully simulate the competition between glutamate and γ-aminobutyric acid (GABA) in a multiplexed transmission process in the human nervous system. To emulate action signals that respond to environmental stimuli in a low-level nerve pathway, a fiber-level neurologically integrated muscular system was constructed by cascading with FNDs and artificial muscle fibers; the system realized unconditioned reflex, even under loads of several Newtons. To emulate the high-level nerve pathway, multiple conductive states of FNDs were used to construct flexible neuromorphic networks; the recognition accuracy for the Fashion MNIST dataset was > 83%, with < 0.1% loss of accuracy even after 100 bending cycles, which represents the most stable recognition result for flexible neuromorphic electronics so far. The presented FNDs provide an excellent basis for the development of human-compatible artificial neurological systems. • We integrated electrochemical graphene artificial synapse on a copper wire as fibrous neuromorphic devices (FNDs) exhibiting multiplexed neurochemical transmission. • Owing to the unique property of tunable charge-carrier polarity, FNDs successfully simulate the competition between glutamate and γ-aminobutyric acid (GABA) in a multiplexed transmission of the biological nerve pathway. • To emulate action signals in a low-level nerve pathway that responds to environmental stimuli, we cascaded artificial muscle fibers and FNDs to construct an all-fiber neurologically integrated muscle system and successfully implemented unconditioned reflex even under loads of several newtons. • To present the implementation of a high-level nerve pathway, FNDs with multiple conductive states were tested in flexible neuromorphic networks; the recognition accuracy for F-MNIST exceeded 83%, within a margin of error of < 0.1% even after 100 bending cycles.