A memristive nanoparticle/organic hybrid synapstor for neuro-inspired computing

A large effort is devoted to the research of new computing paradigms associated to innovative nanotechnologies that should complement and/or propose alternative solutions to the classical Von Neumann/CMOS association. Among various propositions, Spiking Neural Network (SNN) seems a valid candidate. (i) In terms of functions, SNN using relative spike timing for information coding are deemed to be the most effective at taking inspiration from the brain to allow fast and efficient processing of information for complex tasks in recognition or classification. (ii) In terms of technology, SNN may be able to benefit the most from nanodevices, because SNN architectures are intrinsically tolerant to defective devices and Published on line: Dec. 13, 2011 2 performance variability. Here we demonstrate Spike-Timing-Dependent Plasticity (STDP), a basic and primordial learning function in the brain, with a new class of synapstor (synapsetransistor), called Nanoparticle Organic Memory Field Effect Transistor (NOMFET). We show that this learning function is obtained with a simple hybrid material made of the selfassembly of gold nanoparticles and organic semiconductor thin films. Beyond mimicking biological synapses, we also demonstrate how the shape of the applied spikes can tailor the STDP learning function. Moreover, the experiments and modeling show that this synapstor is a memristive device. Finally, these synapstors are successfully coupled with a CMOS platform emulating the preand post-synaptic neurons, and a behavioral macro-model is developed on usual device simulator. 1. Introduction Spike-Timing Dependent Plasticity (STDP) is widely believed today to be one of the fundamental mechanisms of the unsupervised learning in biological neural networks. STDP in biological systems is a refinement of Hebb’s learning rule.[] Grant et al.[], Markram et al.[], Bi and Poo [4] observed STDP in biological synapses. The principle of STDP is to tune the response of a synapse as a function of the preand postsynaptic neurons spiking activity Fig. 1-a. Depending on the correlation or anti-correlation of the spiking events of the preand post-synaptic neurons, the synapse’s weight is reinforced or depressed, respectively. The socalled "STDP function" or "STDP learning window" is defined as the relationship between the change in the synaptic weight or synaptic response versus the relative timing between the preand post-synaptic spikes (Fig. 1-b).[] The implementation of STDP with nanodevices is strongly driven by a bio-inspired approach to enable local and unsupervised learning capability in large artificial SNN in an efficient and robust way. To this end, it is envisioned to use the nanodevices as synapses and to realize the neuron functionality with CMOS. This Published on line: Dec. 13, 2011 3 approach is supported by the fact that the limiting integration factor is really the synapse density, as realistic applications could require as much as 10 to 10 synapses per neuron. Snider [6] proposed an implementation of STDP with nanodevices, where the synapses are realized with a crossbar of memristors [7] and the neurons with a “time-multiplexing CMOS” circuit. Using these two elements, it should be possible to reproduce exactly the “STDP learning window” of a biological synapse (Fig. 1-b). Linares-Barranco et al. simulated the implementation of the STDP function with memristive nanodevices.[] Using a specific shape of the spikes and the non-linearity of the memristor, they showed that the conductivity of the memristor can be tuned depending on the precise timing between the post-synaptic and pre-synaptic spikes. More interestingly, they showed that the shape of the STDP learning window can be tuned by changing the shape of the spike (Fig1-c). We have to emphasize that our aim is to be inspired by the behavior of a biological synapse for neural computation applications (and not to build a model system of the synapse), thus the important point is to reproduce qualitatively the STDP behavior, even if the spike signals applied to the synapstor are not close to the real biological spike. We recently demonstrated that the Nanoparticle-Organic Memory FET (NOMFET) is able to mimic the short-term plasticity (STP) behavior of a spiking biological synapse.[] When a sequence of voltage pulses is applied across the device, the current transmitted by the NOMFET is modulated depending on the frequency of the pulses and the past input activity of the device,[] mimicking the facilitating or depressing behavior of a biological spiking synapse.[] Research on artificial synapse devices mimicking the plasticity of a biological synapse is a burgeoning field. Recently, Jo et al.[] have observed STDP in Ag/Si-based memristor, Lai et al.[] in polymer/Si nanowire transistor, Seo et al.[] in oxide resistive memory, Kuzum et al. in phase-change memory.[] Here, we demonstrate the STDP behavior Published on line: Dec. 13, 2011 4 of the NOMFET. First, we carefully analyze the behavior of this synapstor and show that it can be modelized by the memristor equations.[] Thus, we follow the Linares-Barraco et al. suggestions [8,9 to successfully implement the STDP behavior with the NOMFET. Beyond the demonstration at a single device level, we also demonstrate that the NOMFET can be efficiently coupled with a CMOS platform emulating the preand post-synaptic neurons. Finally, we developed a behavioral macro-model suitable for device/circuit simulations using commercially available simulators (Spectre-Cadence). 2. The NOMFET: a memristive device. The NOMFET is based on a standard bottom gate/bottom source-drain organic transistor with gold nanoparticles (NPs) fixed at the gate dielectric/organic semiconductor (OSC) interface by surface chemistry (see Experimental section, and a detailed material characterization in Ref. []). The STP behavior of the NOMFET is due to the internal charge/discharge dynamics of the NP/OSC system with typical time constants that can be adjusted between 1 to 10 s.[] While we have demonstrated some simple neuro-inspired plasticity for NOMFETs with a channel length L down to 200 nm, and NP diameter of 5 nm, working at a nominal bias of – 3V,[] here for the sake of demonstration, all the experiments are reported for L = 5 μm NOMFETs and 20 nm diameter NPs working at a nominal voltage of -30V, because these devices previously showed the largest plasticity amplitude (i.e. the largest modulation of the NOMFET output current, here analogous to the synaptic weight, by the applied spike sequence).[] The channel width (W) is 1,000 μm for the 5 μm length NOMFE, to maximize the output current, given the relative low mobility of the device (ca. 10 cmVs).[] Optimization of the OSC properties (not done here) will allow reaching a state-to-the art mobility of about 1 cmVs, and will allow reducing the actual width by a factor 10. Published on line: Dec. 13, 2011 5 Further optimization would be the use of high-k dielectric to reach the same output current while downscaling W accordingly. Downscaling the NOMFET channel length to 30 nm (with 5 nm diameter NPs) is possible (we have already demonstrated a 30 nm channel length OFET[]), but such a task would require a hard work for technological optimization, out of the scope of this proof of principle demonstration. The NOMFET is used as a pseudo two-terminal device (Fig. 2-a): the drain (D) and gate (G) electrodes are connected together and used as the input terminal of the device, and the source (S) is used as the output terminal (virtually grounded through the ampmeter). To establish that it works as a memristive device, we write the output current input voltage relation in the NOMFET according to the formalism proposed by Chua,[] and we discuss the significance of the terms in this equation: IDS (t) = G(QNP (t),VDS (t),t)VDS (t) (1) ̇ Q NP (t) = g(QNP (t),VDS (t),t) (2) where G is the conductance of the device that includes the field effect, VDS(t) is the applied signal of time varying shape, and QNP(t) the charges trapped in the NPs. For the NOMFET, QNP(t) is the relevant internal parameter, and its first-order time derivative is given by the g function, which is the "memristive" function that describes how this internal parameter is updated as function of the internal state, the external voltage and time. A non-linear behavior of g is very interesting to implement synaptic plasticity and STDP.[] A g function with a null value between negative and positive threshold voltages and increasing/decreasing parts above/below (respectively) these thresholds has been used to simulate STDP and learning capabilities in memristor-based neuro-inspired circuits.[] To characterize the memristive behavior of the NOMFET, we measure the change of its internal parameter δQNP when voltage signal VDS(t) is a pulse of amplitude VP and duration Published on line: Dec. 13, 2011 6 10 s. This value of 10 s has been fixed in order to maximize the effect of the NP charge. This time is longer than the typical charging/discharging time constants (about 2-3 s)[] for a NOMFET with a channel length of 5 μm and 20 nm NPs used for these experiments. Reducing the width of the charging pulse will give smaller variations of the current, but does not change the conclusions. The output current, before (IInitial) and after (Iafter) the application of the charging pulse, are measured with a short read pulse (100 ms). This pulse is short enough to not modify the charge state of the NPs. Plotting (Iafter – Iinitial)/Iinitial, which is proportional to (Eq. S24, supporting information), versus VP gives a representation of the g-function of the NOMFET. As the current at a given time t depends on the history of the device, we have developed a specific reset protocol (see Experimental section, and Fig. S1, supporting information) that sets the charge state of the NPs to the same before each measurement at different VP. Figure 2-c shows the measured relative variation of the current (red dots) as a function of VP, i.e. the internal memristive-like function of the NOMFET. This function displays the three expected regions similarly to the resistance change in a voltage-controlled memristance:[] (i) For the negative voltage, the NPs are charged with holes, the Coulomb repulsion between the positively charged NPs and the OSC reduces the hole density in the conducting channel, the conductivity of the NOMFET is decreased. (ii) For intermediate voltages (Vth1 < V < Vth2), the effect of the input voltage on the charge state of the NPs is null. The charge state of the NPs cannot be changed. The physical meanings of the two threshold voltages, Vth1 ≈ 0 V and Vth2 ≈ 15 V, are discussed in the supporting information. (iii) For large positive voltages, holes can be detrapped from the NPs, leading to a reverse effect, i.e. an increase in the conductivity of the NOMFET. The memristive g function shown in Fig. 2-c can be calculated using Eqs. S31 (see supporting information) considering the three parts of the experimental curve. For simplicity, we assume Published on line: Dec. 13, 2011 7 the same time constants in Eqs. S31 (τ = τ0 = τ+ = τ= 5 s). This value is in good agreement with experimental values for the NOMFET.[] The blue squares in Fig. 2-c are the fit of this model. Eq. S31 gives two linear relationships for the two branches that fit relatively well our data.