Source-linked AI summary
A memristive nanoparticle/organic hybrid synapstor for neuro-inspired computing
F. Alibart, S. Pleutin, O. Bichler, C. Gamrat, T. Serrano-Gotarredona, B. Linares-Barranco, D. Vuillaume
TL;DR
Spiking neural networks motivate nanodevice synapses that can implement timing-based learning efficiently. This paper develops the NOMFET synapstor, demonstrates STDP and spike-shape control, and integrates it with CMOS-emulated neurons and a behavioral model.
Problem
The paper addresses the need for alternative nanotechnology-based computing architectures supporting timing-dependent, local, and unsupervised learning.
Method
The authors develop a gold-nanoparticle/organic-semiconductor NOMFET synapstor and model its operation as a memristive device.
Results
The NOMFET demonstrates STDP, spike-shape-dependent learning-window control, CMOS-neuron coupling, and a behavioral macro-model.
Takeaways & Limitations
The device provides a qualitative neuromorphic synapse implementation whose correlated and anti-correlated responses resemble biological potentiation and depression.
Takeaways & Limitations
The demonstrated STDP amplitude is weaker than the biological range, and improving the neuro-inspired behavior remains an open technology goal.
Abstract
from arXiv · showhide
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 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 (synapse-transistor), called Nanoparticle Organic Memory Field Effect Transistor (NOMFET). We show that this learning function is obtained with a simple hybrid material made of the self-assembly 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 pre- and post-synaptic neurons, and a behavioral macro-model is developed on usual device simulator.
1. Introduction
The introduction motivates nanodevice-based spiking neural networks and presents the NOMFET as a synapstor implementing STDP. The paper analyzes its memristive behavior, spike-shape dependence, CMOS integration, and simulation model.
- Motivation: STDP adjusts synaptic response according to the relative timing of pre- and post-synaptic spikes.The synaptic response is represented here by device conductance.
- Motivation: Nanodevice synapses paired with CMOS neurons are proposed to support local, unsupervised learning in large spiking neural networks.Realistic applications could require 10^3 to 10^4 synapses per neuron.
- Related work: Prior work showed that memristive devices can tune conductivity with spike timing and modify the STDP learning window through spike shape.These studies motivated using memristive nanodevices for neuromorphic circuits.
- Contribution: The NOMFET is introduced as a nanoparticle-organic synapstor designed to reproduce qualitative STDP behavior rather than biological spike waveforms exactly.The device is based on self-assembled gold nanoparticles and organic semiconductor material.
- Contribution: The paper demonstrates NOMFET STDP, models the device with memristor equations, couples it to CMOS-emulated neurons, and develops a behavioral macro-model.The macro-model is intended for device and circuit simulation.
2. The NOMFET: a memristive device.
This section characterizes the NOMFET as a memristive device whose trapped nanoparticle charge controls conductivity. Its measured voltage response exhibits thresholded regions that are captured qualitatively by a model.
- Device and operation: The NOMFET uses gold nanoparticles at the gate-dielectric/organic-semiconductor interface and is operated as a pseudo two-terminal device.Drain and gate form the input terminal, while the source serves as the output terminal.
- Device limitations: The demonstrated device is physically large because of low mobility, although optimization could reduce its width and channel length.The present proof-of-principle scope excludes the required technological optimization.
- Memristive model: The trapped nanoparticle charge QNP(t) is the internal state variable governing the device’s memristive behavior.The memristive function updates this state according to internal state, external voltage, and time.
- Measured behavior: Negative voltage charges the nanoparticles and decreases conductivity, whereas large positive voltage detrapping increases conductivity.Between the thresholds, the nanoparticle charge state remains unchanged.
- Model validation: A model using voltage-dependent charge/discharge dynamics reproduces the measured memristive curve with two approximately linear branches.The fit assumes equal time constants τ = 5 s, consistent with experimental values.
3. STDP behavior of the NOMFET
The NOMFET reproduces timing-dependent potentiation and depression by combining pre- and post-synaptic spikes. Changing spike shape modulates the STDP learning window, and the model qualitatively matches the experiments.
- STDP response: Pre-synaptic spikes alone leave the NOMFET conductivity unchanged, establishing that isolated pre-synaptic activity does not alter synaptic weight.Ten pre-synaptic spikes were applied in the no-post-spike condition.
- STDP response: Correlated spikes with Δt = +2 s increase conductivity and reinforce the synaptic weight through nanoparticle discharge.The correlated condition used 13 spikes.
- STDP response: Anti-correlated spikes with Δt = −2 s decrease conductivity and depress the synaptic weight as nanoparticles become more charged.The negative effective-voltage contribution dominates in this condition.
- Spike-shape dependence: The STDP learning window is successfully modulated by changing the applied pulse shape from triangular to rectangular.The rectangular-pulse response becomes more squared or rounded than the triangular-pulse response.
- Model comparison: The behavioral model reproduces the experimental STDP curves qualitatively using voltage-dependent time constants from 0.3 to 5 s.These values agree with previously measured multi-time-constant nanoparticle charging and discharging.
- Scope and limitation: The STDP amplitude ranges from −15% to 30%, below the −40% to 100% range reported for biological synapses.The authors identify technology optimization as a route for improvement.
4. Hybrid NOMFET/CMOS system.
The NOMFET synapstors were integrated with a CMOS electronic board that generated pre- and post-synaptic spikes, producing reproducible STDP behavior across devices and initial states. The results indicate that temporal coding can program reliable conductivity changes despite device variability.
- System implementation: A CMOS electronic board generated pre- and post-synaptic spikes and applied them to several NOMFETs.The board was FPGA-driven, remotely controlled by a PC, and directly connected to the synapstors.
- System implementation: The NOMFET/CMOS system produced an STDP function that agreed well with point-by-point probe-station measurements.The comparison used simultaneous measurements from two NOMFETs.
- STDP behavior: The system retained correct STDP behavior when the timing between pre- and post-synaptic spikes was randomized.The randomized-timing measurements were compared with measurements using a linear Δt sequence from -5 to +5 s.
- STDP behavior: STDP behavior remained consistent and reproducible despite differences in device characteristics.The two NOMFETs had a factor-of-10 difference in mean conductivity ratio, while their relative conductivity changes were the same.
- Implications: Temporal coding and natural nanoparticle relaxation make relative conductivity changes reliable even in seemingly unreliable devices.The authors connect this behavior to programming conductivity changes with temporal information rather than absolute device properties.
5. Behavioral macro-model for neuro-inspired circuit simulation
A behavioral macro-model represents the NOMFET through an internal voltage linked to nanoparticle charge and a circuit time constant. Simulations reproduced measured signal behavior, validating the model for simulating STDP-based neuro-inspired circuits.
- Model structure: The physical model represents the diode-connected NOMFET using an internal voltage w(t) proportional to nanoparticle charge.The internal voltage is generated by an R-C circuit driven by a current source, with time constant τ = RC.
- Application: The validated macro-model can be used to simulate neuro-inspired circuits implementing STDP learning with NOMFET devices.The model was implemented in SPECTRE-CADENCE for neuro-inspired circuit simulation.
- Model structure: The behavioral macro-model adapts NOMFET terminal voltages and sensed diode current for implementation in a silicon-device simulator.Voltage scaling accommodates the NOMFET’s operating range of a few tens of volts.
- Simulation: The model was stimulated with a -35 V, 10 s pulse to simulate NOMFET signal evolution.The simulated device was evaluated with VS = 0 while the pulse was applied at VD.
- Simulation: Simulated results showed very good agreement with the measured IDS signal.The fitted time constant was τ = 2.2 s, consistent with experimental NOMFET values reported elsewhere.
6. Discussion and Conclusion
The NOMFET reproduces spike-timing-dependent synaptic behavior through nanoparticle charge dynamics, with responses resembling biological synapses but operating on different timescales. Its speed and biological similarity remain limited by device materials and charge transport.
- Biological comparison: Potentiation for correlated spikes and depression for anti-correlated spikes resemble biological synaptic behavior, although the NOMFET mechanisms differ biologically.The response occurs at different timescales because the two systems have different internal dynamics.
- Outlook: The authors identify room to improve the NOMFET’s neuro-inspired behavior.Earlier devices with a 200 nm channel and 5 nm nanoparticles showed short-term plasticity with approximately 1 s time constants but weaker amplitude.
- Outlook: The discussion connects the synapstors to their future use in neuro-inspired computing circuits and architectures.The supplied passage states this as a future-use direction without specifying a demonstrated system-level application here.
- Device dynamics: The NOMFET’s short-term response is attributed to nanoparticle charge/discharge time constants and low charge mobility in the pentacene/nanoparticle channel.Alkyl chains act as a tunnel barrier, while low charge mobility reduces device functioning speed.
- Improvement routes: Shorter timescales closer to those of biological synapses may be attainable by changing nanoparticle capping molecules or optimizing the organic semiconductor.Suggested routes include more conducting π-conjugated molecules and increased charge-carrier mobility.
7. Experimental
The experimental section describes fabrication, controlled electrical measurement, and reset procedures for nanoparticle/organic synapstors. The devices use functionalized gold nanoparticles, pentacene, and controlled environmental conditions.
- Device fabrication: The devices are fabricated on highly doped p-type silicon covered by 200 nm silicon dioxide, with titanium/gold electrodes and a pentacene film.The pentacene layer is evaporated to a thickness of 35 to 50 nm.
- Controlled environment: Fabrication and electrical measurements are performed under argon or controlled nitrogen atmospheres to limit degradation of the organic materials.The glove-box atmosphere contains less than 1 ppm oxygen and water vapor during relevant procedures.
- Electrical measurements: Input spikes and spike sequences are designed with Matlab and delivered by an arbitrary waveform generator, while output currents are measured with a semiconductor parameter analyzer.The generator is remote controlled by a PC, and measurements use an Agilent 4155C analyzer.
- Reset protocol: The reset protocol applies a decreasing sinusoidal voltage so nanoparticles are alternately charged and discharged with decreasing magnitude.This procedure allows measurements to begin from the same initial condition.
Supporting Information.
The supporting information accompanies the paper on a memristive nanoparticle/organic hybrid synapstor for neuro-inspired computing. It lists the authors and their institutional affiliations.
- Document scope: The supporting information concerns a memristive nanoparticle/organic hybrid synapstor for neuro-inspired computing.This wording also appears in the paper title and subtitle.
- Authorship: The paper lists F. Alibart, S. Pleutin, O. Bichler, and C. Gamrat among its authors.The supplied author list continues with additional coauthors.
- Authorship: Additional listed authors are T. Serrano-Gotarredona, B. Linares-Barranco, and D. Vuillaume.The authors are associated with IEMN, CEA LIST, and IMSE in the supplied affiliations.
- Affiliations: The listed institutions include IEMN in Villeneuve d’Ascq, CEA LIST in Gif-sur-Yvette, and IMSE in Seville.The supplied text provides addresses for these institutions.
Theory
The theory models transport, charge accumulation, nanoparticle trapping, and relaxation to explain the NOMFET’s memristive behavior. It also identifies limits of the simplified dynamics and fits experimental data with voltage-dependent parameters.
- Charge transport: Organic-film conduction is modeled as hopping through localized states whose energies are exponentially distributed.The framework follows the Vissenberg–Matters approach and includes tunneling and thermally assisted tunneling.
- Charge transport: Gate bias forms an accumulation layer while charges are also stored in the nanoparticles, with Coulomb interactions shifting the organic-film site energies.The model tracks holes in the pentacene accumulation layer and nanoparticles through QP(t), QNP(t), and their total QT(t).
- Memristive model: The NOMFET is modeled as a memristive device whose current depends on applied VDS and the internal nanoparticle-charge state QNP(t).The charge state modifies the memristive function and the conductivity, while QNP evolves dynamically under voltage-dependent charging and discharging rates.
- Memristive model: The charge dynamics use first-order kinetic equations in which nanoparticle charging and discharging rates depend on VDS.At fixed VDS, stationarity makes the charge variables time-independent, although their values still depend on VDS.
- Model limitations: At VDS=0, the simplified model predicts exponential relaxation from an initial charge, whereas measured relaxations generally follow power laws.The single-exponential approximation is therefore valid only over a limited observation interval; larger time windows require more sophisticated theory.
- Experimental agreement: VDS=-25V and V+=40V give the best fits for characteristic time constants, and one parameter set reasonably fits the full data set.Approximation 2 reproduces the memristive function and the data shown in the referenced figures.