Fig. 2: Fabrication, characterization and analysis of the analogue, VCM-type memristors used in this study.
From: Actor–critic networks with analogue memristors mimicking reward-based learning
a, Schematic of the material composition and biasing scheme. The active layers consist of HfO2 and a CMO. Our memristors are operated by grounding the TiN bottom electrode (BE) and applying an electrical signal to the W top electrode (TE). A conductive filament (CF) is initially created during a forming step by applying a negative voltage at the TE (Supplementary Fig. 2). b, Scanning electron microscopy image providing an angled view of the vertical VCM stack. The active area of the artificial synapse is given by the overlap of the top and bottom electrodes. c, Focused-ion-beam (FIB) cross-section image cut along the dotted line shown in b. The inset shows the deposited material stack including the active HfO2–CMO bilayer of the VCM memristor. d, Resistance–voltage (R–V) characteristics displaying the resistive switching behaviour of the fabricated devices. They exhibit reproducible switching with gradual and quasi-symmetrical set and reset processes. e, Dynamic measurements displaying the synaptic potentiation and depression curves as a function of the number of applied pulses. Data from ten distinct cycles, along with the mean, are shown. Identical pulse trains were used to perform this measurement: 200 set pulses at 2.5 V with a 1.5-μs pulse width for potentiation and 200 reset pulses at –2.7 V with a 10-μs pulse width for depression. The pulse parameters (amplitude and duration) were specifically chosen to achieve a good compromise between the linearity of potentiation/depression and the noise in weight updates. f, Histograms showing the deviations of both potentiation and depression measurements from their respective means for 27 distinct devices.
