Researchers at Dongguk University in South Korea have developed a battery-free, flexible graphene transistor that mimics biological synapses and can recognize human activity without any external power source.
A self-powered graphene-channel transistor, driven entirely by triboelectric nanogenerators, mimics biological synaptic behavior and recognizes human activity from touch alone.
The device is a graphene-channel ion-gel-gated transistor (g-IGT), fabricated on a flexible PET substrate, with monolayer CVD graphene as the channel and an ion-gel electrolyte (based on the ionic liquid EMIM-TFSI) serving as the gate dielectric. Two triboelectric nanogenerators (TENGs), built from a poly(vinylidene fluoride-co-trifluoroethylene) triboelectric layer, are coupled to the transistor: one connected to the gate supplies pre-synaptic spikes, and one connected to the drain supplies post-synaptic spikes. As the TENGs undergo repeated contact and separation from touch or mechanical motion, they generate voltage pulses that drive ion migration within the gel and modulate the graphene channel's conductance - reproducing synaptic signaling entirely from harvested mechanical energy, with no external power supply or battery.
The researchers showed the device reproduces a hierarchy of memory behaviors seen in biological neural networks: sensory memory with a decay time of about 70 milliseconds, short-term memory decaying over 0.2–0.45 seconds, and - with repeated stimulation - a transition toward long-term memory with decay times exceeding 2 seconds. The device also exhibited spike-rate-dependent plasticity, where synaptic strength shifts between potentiation and depression depending on the relative firing rates at the gate and drain, and this behavior remained stable under 1.1 MPa of bending stress.
To test practical use, the team incorporated the device's measured synaptic response into a single-layer perceptron trained to classify six human activities - walking, sitting, standing, lying, walking upstairs, and walking downstairs - using the public UCI Human Activity Recognition dataset. The self-powered, TENG-driven device reached 88.05% classification accuracy under bending, within about 1 percentage point of an ideal software model, and the system maintained better than 75% accuracy even under a high level of simulated signal noise.
The researchers point to self-powered wearable health monitors, electronic skin, smart prosthetics, and human-machine interfaces as potential applications.
