Fig. 3: Non-volatile synaptic behaviour and NNs.

a, Schematic of the cv-OECT acting as a non-volatile synapse. b, Non-volatile conduction changes in the cv-OECT as a function of gate pulse amplitude (representative data from n = 15 continuous pulses, presented as mean values ± s.d.). The pulse greater than |−0.8| V can switch the cv-OECT to the non-volatile mode. Signals are filtered to eliminate the instrument noise and volatile spikes. c, Cyclic transfer curves of cv-OECT with polarizable gate under HGP. The centrosymmetric drain and gate currents with obvious hysteresis demonstrate the non-volatile nature. d, State retention of six specific analogue states (with the gate grounded, which are similar hereafter). Each state can be maintained for longer than 10,000 s with low drift. e, Cyclic LTP under current control (2,000 pulses, ±200 nA, 40 ms). Three reproducible LTPs with linear, symmetrical programming and one-to-one correspondence are shown (top). f, LTP of cv-OECT under voltage control. The blue/pink boxes display 1,024/50 analogue states with low switching stochasticity. g, Architecture of the homogeneous SNN based on cv-OECT. The inset shows a homogeneous 1T1R unit that contains two identical cv-OECTs. Here, SL, WL and BL denote sourceline, wordline and bitline, respectively. h, Plot of the conductance change in STDP and time delay (Δt) between the pre-/post-spike (representative data from n = 5 continuous spikes, presented as mean values ± s.d.). The inset shows the sequence and amplitude of the pulses used for the training of SNN. i, Recognition accuracies of the handwritten images from the MNIST dataset using the experimental non-ideal factors of cv-OECTs and RRAMs.