Inkjet-printed niobium tungsten oxide thin-film memristors for neuromorphic computing

In this work, a Ag/Nb18W16O93 (NWO)/FTO memristor array was fabricated mainly using inkjet printing technology. Resistive switching memory characteristics including a memory window of 102, the retention of 104 s and endurance of 104 was demonstrated. Other than that, synaptic behaviors such as paired-pulse facilitation (PPF), long-term potentiation/depression (LTP/LTD), and spike-time dependent plasticity (STDP) were also demonstrated, showing that the cross-bar NWO-based devices can mimic the behaviors of biological synapses. Based on an array of NWO-based memristors, the LeNet-5 model and an improved version of the traditional nonlinear stochastic gradient descent (SGD) algorithm were used for network training. As a result, a recognition accuracy of the Mixed National Institute of Standards and Technology (MNIST) dataset up to 95.7 % was achieved. These results suggest that the Ag/NWO/FTO memristor array can be potentially used for neuromorphic computing.