Learning through ferroelectric domain dynamics in solid-state synapses

Abstract

In the brain, learning is achieved through the ability of synapses to reconfigure the strength by which they connect neurons (synaptic plasticity). In promising solid-state synapses called memristors, conductance can be finely tuned by voltage pulses and set to evolve according to a biological learning rule called spike-timing-dependent plasticity (STDP). Future neuromorphic architectures will comprise billions of such nanosynapses, which require a clear understanding of the physical mechanisms responsible for plasticity. Here we report on synapses based on ferroelectric tunnel junctions and show that STDP can be harnessed from inhomogeneous polarization switching. Through combined scanning probe imaging, electrical transport and atomic-scale molecular dynamics, we demonstrate that conductance variations can be modelled by the nucleation-dominated reversal of domains. Based on this physical model, our simulations show that arrays of ferroelectric nanosynapses can autonomously learn to recognize patterns in a predictable way, opening the path towards unsupervised learning in spiking neural networks.

ARTICLE

Received 17 Oct 2016 | Accepted 26 Jan 2017 | Published 3 Apr 2017

Sren Boyn1,w, Julie Grollier1, Gwendal Lecerf2, Bin Xu3, Nicolas Locatelli4, Stphane Fusil1, Stphanie Girod1,w, Ccile Carrtro1, Karin Garcia1, Stphane Xavier5, Jean Tomas2, Laurent Bellaiche3, Manuel Bibes1,Agns Barthlmy1, Sylvain Saghi2 & Vincent Garcia1

In the brain, learning is achieved through the ability of synapses to recongure the strength by which they connect neurons (synaptic plasticity). In promising solid-state synapses called memristors, conductance can be nely tuned by voltage pulses and set to evolve according to a biological learning rule called spike-timing-dependent plasticity (STDP). Future neuromorphic architectures will comprise billions of such nanosynapses, which require a clear understanding of the physical mechanisms responsible for plasticity. Here we report on synapses based on ferroelectric tunnel junctions and show that STDP can be harnessed from inhomogeneous polarization switching. Through combined scanning probe imaging, electrical transport and atomic-scale molecular dynamics, we demonstrate that conductance variations can be modelled by the nucleation-dominated reversal of domains. Based on this physical model, our simulations show that arrays of ferroelectric nanosynapses can autonomously learn to recognize patterns in a predictable way, opening the path towards unsupervised learning in spiking neural networks.

1 Unit Mixte de Physique, CNRS, Thales, Univ. Paris Sud, Universit Paris-Saclay, Palaiseau 91767, France. 2 University of Bordeaux, IMS, UMR 5218, Talence F-33405, France. 3 Department of Physics and Institute for Nanoscience and Engineering, University of Arkansas Fayetteville, Arkansas 72701, USA. 4 Centre de Nanosciences et de Nanotechnologies, CNRS, Univ. Paris Sud, Universit Paris-Saclay, C2NOrsay, Orsay cedex 91405, France. 5 Thales Research and Technology, 1 Avenue Augustin Fresnel, Campus de IEcole Polytechnique, Palaiseau 91767, France. w Present addresses: Electrochemical Materials, ETH

Zurich, 8092 Zurich, Switzerland (S.B.); Materials Research and Technology Department, Luxembourg Institute of Science and Technology (LIST), 41 rue du Brill, L-4422 Belvaux, Luxembourg (S.G.). Correspondence and requests for materials should be addressed to S.B. (email: mailto:[email protected]

Web End [email protected] ) or to V.G. (email: mailto:[email protected]

Web End [email protected] ).

NATURE COMMUNICATIONS | 8:14736 | DOI: 10.1038/ncomms14736 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 1

DOI: 10.1038/ncomms14736 OPEN

Learning through ferroelectric domain dynamics in solid-state synapses

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14736

Cortical information ows from neuron to neuron through synapses of variable connection strength. The overall distribution of the synaptic strengths provides the neural

network with memory, while learning is achieved through the synapses reconguration (that is, plasticity)1. Several mechanisms regulating the evolution of the synaptic strengths have been proposed2. A particularly promising one is spike-timing-dependent plasticity3 (STDP) through which the synaptic strengths evolve depending on the timing and causality of electrical signals from neighbouring neurons4 (sketch in Fig. 1a). As observed in biological systems5, STDP enables learning without any external control on the synaptic strengths or any previous knowledge of the information to be processed. This makes STDP the basis for autonomous, unsupervised learning6. The implementation of STDP in articial neural networks thus emerges as a crucial milestone towards the realization of self-adaptive electronic architectures.

An electronic equivalent of the synapse for articial neural networks is the memristor7, a nanoscale device whose resistance depends on the history of electrical signals it was previously subjected to (ref. 8). Memristors thus exhibit plasticity and their conductance can emulate synaptic strength, so that a low resistance corresponds to a strong synaptic connection and a high resistance corresponds to a weak synaptic connection, respectively. Most memristors operate through the motion of ions or atoms in binary oxides (TiO2 (refs 9,10), Ta2O5 (ref. 11) and so on), Ag-Si/Ag-S nanocomposites12 or phase change materials13. Recently, STDP was demonstrated in individual devices based on such memristor technologies, conrming the potential of memristors for autonomous learning1320. However, the connection between the STDP process and the physical mechanisms underlying resistance changes in these memristors is unclear7. In addition, continual variations of the memristor

conductance are required for learning new features under an incessant information ow. Here we show that resorting to purely electronic memristors with a high endurance, operating on well-established physical principles is a decisive asset for the future implementation of unsupervised learning21 in high-density memristive crossbar arrays22.

ResultsFerroelectric synapses. We work with purely electronic memristors based on ferroelectric tunnel junctions (FTJs), in which an ultrathin ferroelectric lm is sandwiched between two electrodes23,24. In such devices, sketched in Fig. 1b, the junction resistance sensitively depends on the relative fraction of ferroelectric domains with polarization pointing towards one or the other electrode25. Applying voltage pulses to the device modies the domain population, thereby inducing resistance changes25. Furthermore, supertetragonal BiFeO3 (BFO) tunnel barriers combined with (Ca,Ce)MnO3 (CCMO) bottom and Co top electrodes give rise to OFF/ON resistance ratios up to 104 (ref. 26) paired with high endurance and operation speed27.

Figure 1c shows the dependence of the junction resistance with the amplitude of 100 ns voltage pulses in a Co/BFO/CCMO junction (Methods). In this hysteresis cycle, one clearly notes the existence of voltage thresholds V th V th

beyond which switching between low- and high- (high- and low-) resistance states occurs. The existence of such well-dened voltage thresholds (associated with the coercivity of the ferroelectric) makes it possible to implement STDP28 in these FTJs. According to STDP, if the pre-neuron spikes just before the post-neuronindicating a causal relationshipthe synapse should be strengthened whereas if the pre-neuron spikes just after the post-neuron indicating an anticausal relationshipthe synapse should be

a

b

PotentialPotential

Vpre

Time

Pre-neuron

Post-neuron

Time

Pt/Co

Synapse

Vpost

BFO

Time

Time

CCMO

YAO

t

c

d

2

V

107

1

Resistance ()

G(S)

0

106

1

V+th

2

t > 0

t

2 1 0 1 2

600 400 200 0 200 400 600

Voltage pulse amplitude (V)

t (ns)

Figure 1 | Articial synapses based on FTJs. (a) Sketch of pre- and post-neurons connected by a synapse. The synaptic transmission is modulated by the causality (Dt) of neuron spikes. (b) Sketch of the ferroelectric memristor where a ferroelectric tunnel barrier of BiFeO3 (BFO) is sandwiched between a bottom electrode of (Ca,Ce)MnO3 (CCMO) and a top submicron pillar of Pt/Co. YAO stands for YAlO3. (c) Single-pulse hysteresis loop of the ferroelectric memristor displaying clear voltage thresholds (V th and V th ). (d) Measurements of STDP in the ferroelectric memristor. Modulation of the device conductance (DG) as a function of the delay (Dt) between pre- and post-synaptic spikes. Seven data sets were collected on the same device showing the reproducibility of the effect. The total length of each pre- and post-synaptic spike is 600 ns.

2 NATURE COMMUNICATIONS | 8:14736 | DOI: 10.1038/ncomms14736 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications

NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14736 ARTICLE

weakened. We emulate the spikes from pre- and post-neurons (sketched in Fig. 1a) by the waveforms shown in Fig. 1b: rectangular voltage pulses followed by smooth slopes of opposite polarity. Importantly, the voltage never exceeds Vth, so that a

single spike cannot induce a change in resistance.

When both pre- and post-neuron spikes reach the memristor with a delay Dt, their superposition produces the waveforms (Vpre Vpost) displayed in the inset of Fig. 1d. The resulting

combined waveform transitorily exceeds the threshold voltage, leading to an increase (DG40, synapse strengthening) or a decrease (DGo0, synapse weakening) of the FTJ conductance (G), depending on the sign of Dt. As can be seen from the experimental STDP curve in Fig. 1d, only closely timed spikes produce a conductance change whereas long delays leave the device unchanged.

Domain dynamics probed by tunnelling. Modelling the shape of the STDP curve requires understanding the physical process underlying the time-dependent variation of the FTJ conductance while the waveform is applied. For this purpose, we image the pseudo real-time evolution of the ferroelectric domain conguration by means of stroboscopic piezoresponse force microscopy (PFM)29, while simultaneously measuring the electrical properties of the device. Figure 2a shows the PFM phase and amplitude signals after cumulative pulses of constant amplitude (1 V) from the ON to the OFF state. The evolution of the phase images in stroboscopic PFM reveals the gradual reversal of the polarization from up (dark domains) to down (bright domains) and is reminiscent of the polarization reversal with increasing voltages previously observed by PFM26. The weak amplitude signal during reversal (for example, stages 2 and 3) indicates that the ferroelectric system is split into many small (t10 nm) domains and that polarization switching is governed by the inhomogeneous nucleation of new domains rather than by the expansion of existing ones. We complemented these experiments by molecular dynamics (MD) simulations (Methods) on defect-free supertetragonal BFO lms. These computations also yield an inhomogeneous character for polarization switching (Supplementary Fig. 1), therefore implying that this process is intrinsic in nature, that is, one does not require the presence of defects to obtain an inhomogeneous switching in supertetragonal BFO. Such ndings contrast with the case of bulk-like BFO (rhombohedral phase) for which a homogeneous switching was recently predicted30 as a result of large oxygen octahedra tilts that provide a specic homogeneous path for polarization reversal.

We extract the normalized reversed area S from the phase images and plot it as a function of the cumulated pulse duration in Fig. 2b (black squares). Owing to the direct link between the junction resistance R and this normalized reversed area S (well described by a simple model of parallel resistances, 1=RG1 S 1=RON S 1=ROFF, ref. 25), one can also

extract S from measurements of the junction resistance after consecutive pulses of 1 V (green squares in Fig. 2b; note the excellent agreement between the PFM and transport data). Figure 2c shows transport data sets at different pulse amplitudes. They follow a systematic trend where, for a given cumulated pulse duration, the switched area is larger under higher voltages.

In ferroelectrics, inhomogeneous polarization switching can be described by a nucleation-limited model, which considers that the ferroelectric lm is composed of different zones with independent switching kinetics31,32. Assuming for each voltage V a broad Lorentzian distribution of the logarithm of nucleation times with width G(V) and centred at log(tmean(V))the normalized

reversed area S can be approximated as a function of time t and

applied voltage V (ref. 31):

S t; V

1

2

1p arc tan

log tmeanV

log t

GV

; 1

where the index relates to the sign of the applied voltage. Fits obtained with this expression accurately reproduce the experimental data of ferroelectric switching as a function of time and voltage (black lines in Fig. 2c). Figure 2d displays several representative distributions of switching times and illustrates the main trends. For larger voltages, switching occurs earlier and in a narrower time window (decrease of tmean and G), in agreement

with previous results obtained on thick ferroelectric capacitors32. As shown in the inset of Fig. 2d, MD simulations performed at10 K conrm the relevance of equation (1) (Methods and Supplementary Fig. 2) to characterize the switching process as well as the measured trends of tmean and G with the magnitude of

the electric eld. Figure 2e shows that the evolution of the switching time (tmean) as a function of the inverse electric eld, extracted from transport measurements (Fig. 2c), follows the characteristic Merzs law32 for ferroelectric switching (tmeanpexp(a/E)) where the activation eld a is of the order of3.0 V nm 1. Due to the idealized nature of BFO during MD simulations (no interface, no defects, no tunnel current), the timescale for polarization switching is much shorter than in experiments, while the electric eld is larger. Nevertheless, Merzs law can be applied to the MD simulation results (inset of Fig. 2e) and indicates an activation eld of 2.4 V nm 1, that is, in the same range as for experimental results. These simulation results strongly suggest that the experimentally observed inhomogeneous polarization switching in ultrathin lms of BFO has an intrinsic origin.

Predictive modelling of synaptic learning. Because of the direct relationship between S and R in these devices, the accurate description of ferroelectric switching by the nucleation-limited model can be further extended to the modelling of resistance changes as a function of voltage amplitude. Figure 3a shows nested resistance hysteresis loops as a function of voltage pulse amplitude, characteristic of memristors. The STDP curve of the same device is displayed in Fig. 3b. Using equation (1) in combination with the parallel resistance model (Methods and Supplementary Fig. 3), we simultaneously t the resistance versus voltage cycles and the STDP curve. Both resistance changes can be accurately replicated (solid lines in Fig. 3a,b) using the same set of parameters tmean(V) and G(V).

This full description for each memristor device makes it possible to predict the conductance changes driving STDP learning in ferroelectric synapses. In Fig. 3bd, we apply different voltage waveforms to our memristors to emulate various types of pre- and post-neuron activities3. This procedure allows the generation of biologically realistic2, though accelerated (Fig. 3b,c), or articially designed (Fig. 3d) STDP learning curves. Using our model with the parameters extracted previously, we can now predict the conductance changes for these specic types of STDP. Figure 3c,d shows the excellent agreement between these predictions and the measured conductance variations associated with different STDP waveforms.

DiscussionWe now use this physical model to simulate unsupervised learning in a spiking neural network with ferroelectric synapses. These simulations serve as a test bench to investigate the inuence of the STDP waveform shape on the ability of the network to recognize patterns in images (here the horizontal, diagonal, and vertical bars labelled A, B and C, as shown in the

NATURE COMMUNICATIONS | 8:14736 | DOI: 10.1038/ncomms14736 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 3

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14736

a

c

ON

1 2 3 4 5 6

360

Max.

Min.

1.0

PFM images

AmplitudePhase

0.8

0

reversed area

b

Normalized reversed area

0.6

Normalized

1.0

4

5

6

3

0.4

0.5

2

1

Transport data PFM data

0.0

ON

0.2

30

ON

20

1

0.0

G(S)

10

2

3

0

4

5

6

10 1

0

107 106 105 104 103 102

Cumulated pulse duration (s)

d e

101

3

Distribution amplitude (a.u.)

103

2

t mean(s)

105

107

1

109

10

0 1

0

1011

108 106 104 102 100

5 10

Switching time (s)

1/E (nm V1)

Figure 2 | A memristor governed by nucleation-limited ferroelectric domain switching. (a) Evolution of the PFM phase and amplitude signals of a ferroelectric memristor under cumulated pulses of 1 V and 100 ns. Cumulative pulses induce a progressive switching with multiple nucleation areas and limited propagation of ferroelectric domains from up (dark phase) to down (bright phase) polarization. The scale bar is 50 nm. (b, top) Normalized switched area as a function of cumulated pulse time calculated from time-dependent transport measurements of a ferroelectric memristor. The black squares indicate the normalized switched area obtained from the PFM measurements in a. (bottom) Corresponding conductance (G) evolution measured as a function of cumulated pulse time. (c) Normalized switched area as a function of cumulated pulse time calculated from time-dependent transport measurements of a ferroelectric memristor at different pulse amplitudes. The black lines are t results from the nucleation-limited switching model. (d) Examples of Lorentzian distributions of switching times extracted from the ts in c at different pulse amplitudes and (inset) from the MD simulations (Supplementary Fig. 2). (e) Evolution of the switching time (tmean) as a function of the inverse of the electric eld (1/E) obtained from ts of the transport data in c and (inset) from MD simulations.

1

a

b

c

d

1

1

1

Data Fit

V(V)

108

0

0

0

Time (ns)

V(V)

1

Time (ns)

V(V)

1

Time (ns)

2.0

Resistance ()

2

1.0

Prediction

107

1.5

1

0.5

G(S)

G(S)

1.0

0

G(S)

0.0

0.5

106

1

0.5

0.0

2

1.0

2 1 0 1 2

600 300 0 300 600

0.5 600 300 0 300 600

600 300 0 300 600

Voltage pulse amplitude (V)

t (ns)

t (ns)

t (ns)

Figure 3 | Predicting STDP learning with ferroelectric synapses. (a) Multiple hysteresis loops of a ferroelectric memristor showing a clear dependence of resistance switching with the maximum pulse amplitudes. (bd) Examples of STDP learning curves of different shapes. The pre- and post-synaptic spikes and conductance variations are shown in the top and bottom panels, respectively. For each device, simultaneous ts of data in a,b (solid lines) using equation (1) allow the prediction of new learning curves in c,d (orange lines).

4 NATURE COMMUNICATIONS | 8:14736 | DOI: 10.1038/ncomms14736 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications

NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14736 ARTICLE

a b

A

B

C

1.0

Inputs

0.9

0.8

Recognition rate

0.7

0.6

Outputs

0.5

0.4

0.3

1 10 100

Number of image presentations

c

d

1.0

Time (ns)

600

0

V(V)

0.0

0.5

0.2

1.0

0.5

V(V)

0.0

0

600

Output neuron spikes

0.8

Recognition rate

0.6

#3

0.4

0.2

#5

0.0

Input

image

0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4

Post spike amplitude A (V)

Figure 4 | Unsupervised learning with ferroelectric synapses. (a) Simulated spiking neural network comprising nine input neurons connected to ve output neurons by an array of ferroelectric memristors. The inputs are noisy images of the patterns to recognize: horizontal (A), diagonal (B) and vertical(C) bars in 3 3 pixel images. (b) Recognition rate as a function of the number of presented images for different noise levels. The coloured images are

conductance maps of the memristors in each line and show their evolution for a noise level of 0.3 (blue line). (c) Behaviour of the network after successful learning. Neurons 1, 2 and 4 emit spikes when inputs C, B and A are presented, respectively. (d) Evolution of the nal recognition rate as a function of the amplitude of the post-synaptic spike. Error bars represent s.d.

inset of Fig. 4a). The simulated network is built around a crossbar of 9 5 ferroelectric memristors (Fig. 4a). Each of the nine

spiking input neurons codes for one pixel of noisy images that contain one out of the three patterns to recognize (Methods). The output neurons integrate the input signals owing through the memristors along each row and re when a threshold is reached. Figure 4b shows that the recognition rate of the network increases with the number of image presentations, reaching 100% for low noise levels and almost 80% for high (of the order of the input amplitude) noise levels. The evolution of the conductance of the nine memristors in each row is shown in the inset of Fig. 4b as coloured 3 3 pixel images. As the network is learning, the

conductance images in rows 1, 2 and 4 each converge toward one out of the three input patterns, whereas the memristor conductances in rows 3 and 5 remain random. After learning, as illustrated in Fig. 4c, three of the output neurons (neurons 1, 2 and 4 in this example) have specialized to each of the three input patterns and re when the corresponding input is presented. Our simulations reveal that successful unsupervised learning is highly dependent on the exact shape of the pre- and post-neuron spike waveforms. Figure 4d shows that, for the case presented before, the recognition rate reaches 100% for well-chosen post-neuron spike amplitudes in the range of 0.820.98 V. However, when the amplitude of the post-neuron spike waveform is only slightly lower or higher, the recognition rate rapidly drops. Our simulations therefore emphasize the importance of a precise knowledge of the memristor dynamics, and therefore of its accurate description on the basis of a physical model.

In summary, we have established that STDP can be harnessed from intrinsically inhomogeneous polarization switching in ferroelectric memristors. Combining time-dependent transport measurements, ferroelectric domain imaging, and effective-Hamiltonian-based atomistic MD simulations, we show that the ferroelectric switching underlying resistive changes in these devices can be described by a well-established nucleation-limited model. Using this physical model, we can reliably predict the conductance evolution of ferroelectric synapses with varying neural inputs. These results pave the way toward low-power hardware implementations of billions of reliable and predictable articial synapses33 (such as deep neural networks34) in future brain-inspired computers.

Methods

Sample fabrication. The BiFeO3 (4.6 nm)/Ca0.96Ce0.04MnO3 (20 nm) thin lms are grown on YAlO3 (YAO) (001) substrates by pulsed laser deposition using a Nd:YAG laser (details of the growth can be found in ref. 26). The thin lms are epitaxial and fully strained by the substrates as characterized by X-ray diffraction experiments26 and scanning transmission electron microscopy on cross-section samples35. The high compressive strain imposed by the YAO substrate on BiFeO3 (BFO) stabilizes the supertetragonal polymorph of BFO with giant tetragonality and a potentially large polarization pointing initially towards the bottom electrode of Ca0.96Ce0.04MnO3 (CCMO)35.

FTJs are then fabricated by dening Pt (10 90 nm)/Co (5 10 nm) nanopillar

electrodes (with diameters of 180500 nm) on top of the unpatterned BFO/CCMO heterostructure, combining conventional electron-beam lithography and lift-off processes. For experiments involving electron transport and ferroelectric domain imaging experiments (as in Fig. 2), arrays of nanopillars are electrically contacted with a conducting atomic force microscopy tip26. For all other electronic transport experiments presented in the manuscript, fully patterned junctions are dened by a four-step lithography process following the denition of these nanopillars (details

NATURE COMMUNICATIONS | 8:14736 | DOI: 10.1038/ncomms14736 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 5

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14736

of the fabrication process can be found in ref. 27). Such fully patterned junctions are macroscopically connected by standard radio frequency (RF) probes.

Physical measurements. Combined transport and PFM measurements are performed using solid-state FTJs with diameters of 180 nm, as detailed in ref. 26. These Pt/Co/BFO/CCMO junctions are connected electrically by a conductive atomic force microscopy tip to allow resistance measurements under constant dc voltage ( 0.1 V, time constant 100 ms) and the application of 100-ns write pulses in

combination with subsequent PFM imaging (Vac 0.6 V, Vdc 0.17 V,

f 14 kHz). The bias voltage is applied to the tip while the bottom electrode is

grounded.

The resistancevoltage cycles and STDP curves were obtained on fully processed FTJs with diameters of 400 and 500 nm, described in ref. 27. The resistance of the junctions was measured after each pulse at a low bias of 200 mV

(time constant 100 ms). To obtain the STDP curve, we always initialize the junction to the ON state for Dto0 and to an intermediate resistance state between ON and

OFF for Dt40. The exact values of initial resistance states in each gure areFigs 1d and 3b: Rinit(Dto0) 6 105 O, Rinit(Dt40) 1.2 107 O,

Fig. 3c: Rinit 1.2 107 O,

Fig. 3d: Rinit(Dto0) 6 105 O, Rinit(Dt40) 1.2 107 O.

Molecular dynamics simulations. The switching process under dc electric elds is simulated using a recently developed scheme36, in which the total energy is provided by the rst-principles-based effective Hamiltonian method of refs 37,38 and inserted into an original MD technique treating on the same footing dynamics of ions and magnetic degrees of freedom. A periodic 48 48 6 supercell (69,120

atoms) of the supertetragonal BFO (P4mm space group) is adopted and the electric eld is constantly applied along the pseudocubic 00 1

direction (opposite to the

initial polarization direction). The temperature is kept at 10 K so as to minimize thermal uctuations and the time step is set to 0.5 fs. Supplementary Fig. 1a shows polarization patterns at different times under a dc electric eld of 2.5 V nm 1, which illustrate the (intrinsic) inhomogeneous switching of polarization in bulk-like supertetragonal BFO. The switched area as function of time is plotted in Supplementary Fig. 1b. Note that the reversed area is dened as any polarization that has a magnitude larger than 7 mC cm 2 (it reaches 100 mC cm 2 when being fully switched) in the 00 1

direction. Simulation results of polarization switching

under different electric elds are summarized in Supplementary Fig. 2. Fits from equation (1) are represented as black lines. Lorentzian switching distributions resulting from the ts are displayed in the inset of Fig. 2d and the switching times are plotted in the inset of Fig. 2e.

Domain dynamics model. The nucleation-limited model as expressed by equation (1) describes the full reversal of the ferroelectric domains starting from a uniform state under constant voltage amplitude. To describe the ferroelectric switching dynamics under the inuence of arbitrary voltage waveforms, the waveform is numerically divided in short segments of width Dt (Supplementary

Fig. 3a). For sufciently small Dt, we can then assume a constant applied voltage Vi.

As the reversal dynamics depends on the present domain conguration and therefore on the value Si 1 before the segment i, we compute the time offset toffseti

corresponding to a voltage Vi and can only then apply equation (1) to calculate the value Si S toffseti Dt; Vi

(Supplementary Fig. 3b). Owing to the highly nonlinear reversal dynamics of FTJs where high voltage amplitudes govern the reversal process, the value of Dt has to be chosen small enough not to smear out these maximal amplitudes. For the results presented here, we chose Dtr20 ns.

A simple yet rigorous model connects the electrical resistance R with the ferroelectric domain conguration characterized by the normalized area of down domains S. Assuming parallel conduction through areas with ferroelectric down domains and areas with up domains allows to calculate one from another using the relation 1=R 1 S 1=RON S 1=ROFF, where RON and ROFF are the

resistance values in the low- and high-resistance states, respectively.

Spiking neural network simulations. The simulated spiking neural network consists of nine input neurons fully connected to ve output neurons by a crossbar of 9 5 identical ferroelectric memristors, as depicted in Fig. 4a. Inputs are con

secutively presented to the network in the form of grey-level images of 3 3 pixels

composed of one of the three patterns to recognize (horizontal, diagonal, and vertical bars labelled A, B and C in the inset of Fig. 4a) and an additional uniformly distributed random noise of the specied amplitude. Subsequently, the values of each image are normalized to the interval [0, 1]. Each individual pixel corresponds to one input neuron that integrates the respective input value over time (using a xed time step) and spikes once it reaches a certain threshold Qth. This imple

mentation effectively results in a frequency coding of the input pixel values. The spikes from each input neuron are transmitted to all output neurons (as in a fully connected network) that integrate them weighted by the corresponding connection strength, that is, by the normalized conductance value of the respective FTJ. Once an output neuron (#1 to #5 in Fig. 4) reaches the threshold value Qth, it spikes on its part and generates an output spike. Lateral inhibition then resets all output neurons. We use the rst output spike after the presentation of a new input image to determine if the network has successfully recognized the pattern (A, B or C as

shown in Fig. 4a) of the input image. As for unsupervised learning the labels of the output neurons are not known beforehand, the labelling is done algorithmically for each repetition of the simulation. We dene a successful learning as when each pattern is immediately recognized after 200 input image presentations. The recognition rates shown in Fig. 4b,d are calculated as the mean of 100 simulation runs for each parameter set. At the beginning of each simulation run, the memristor conductances are initialized to random values between those of their ON and OFF states. Variations between the runs using identical parameters are therefore due to the randomized initial conductance values of the synapses as well as the random noise that is added to the input patterns.

Data availability. The data that support the ndings of this study are available from the corresponding authors upon request.

References

1. Hebb, D. O. The Organization of Behavior (Wiley & Sons, 1949).2. Abbott, L. F. & Nelson, S. B. Synaptic plasticity: taming the beast. Nat. Neurosci.

3, 11781183 (2000).

3. Zamarreno-Ramos, C. et al. On spike-timing-dependent-plasticity, memristive devices, and building a self-learning visual cortex. Front. Neurosci. 5, 26 (2011).

4. Markram, H., Lbke, J., Frotscher, M. & Sakmann, B. Regulation of synaptic efcacy by coincidence of postsynaptic APs and EPSPs. Science 275, 213215 (1997).

5. Bi, G. Q. & Poo, M. M. Synaptic modications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J. Neurosci. 18, 1046410472 (1998).

6. Masquelier, T., Guyonneau, R. & Thorpe, S. J. Competitive STDP-based spike pattern learning. Neural Comput. 21, 12591276 (2009).

7. Yang, J. J., Strukov, D. B. & Stewart, D. R. Memristive devices for computing. Nat. Nanotechnol. 8, 1324 (2013).

8. Chua, L. Memristor-the missing circuit element. IEEE Trans. Circuit Theory 18, 507519 (1971).

9. Strukov, D. B., Snider, G. S., Stewart, D. R. & Williams, R. S. The missing memristor found. Nature 453, 8083 (2008).

10. Szot, K. et al. TiO2-a prototypical memristive material. Nanotechnology 22,

254001 (2011).

11. Lee, M.-J. et al. A fast, high-endurance and scalable non-volatile memory device made from asymmetric Ta2O(5 x)/TaO(2 x) bilayer structures. Nat. Mater. 10,

625630 (2011).12. Jo, S. H. & Lu, W. CMOS compatible nanoscale nonvolatile resistance switching memory. Nano Lett. 8, 392397 (2008).

13. Kuzum, D., Jeyasingh, R. G. D., Lee, B. & Wong, H. S. P. Nanoelectronic programmable synapses based on phase change materials for brain-inspired computing. Nano Lett. 12, 21792186 (2012).

14. Jo, S. H. et al. Nanoscale memristor device as synapse in neuromorphic systems. Nano Lett. 10, 12971301 (2010).

15. Seo, K. et al. Analog memory and spike-timing-dependent plasticity characteristics of a nanoscale titanium oxide bilayer resistive switching device. Nanotechnol. 22, 254023 (2011).

16. Alibart, F. et al. A memristive nanoparticle/organic hybrid synapstor for neuroinspired computing. Adv. Funct. Mater. 22, 609616 (2012).

17. Suri, M. et al. Bio-inspired stochastic computing using binary CBRAM synapses. IEEE Trans. Electron. Devices 60, 24022409 (2013).

18. Li, Y. et al. Ultrafast synaptic events in a chalcogenide memristor. Sci. Rep. 3, 1619 (2013).

19. Mandal, S., El-Amin, A., Alexander, K., Rajendran, B. & Jha, R. Novel synaptic memory device for neuromorphic computing. Sci. Rep. 4, 5333 (2014).

20. Wang, Y.-F., Lin, Y.-C., Wang, I.-T., Lin, T.-P. & Hou, T.-H. Characterization and modeling of nonlamentary Ta/TaOx/TiO2/Ti analog synaptic device. Sci.

Rep. 5, 10150 (2015).21. Choi, S., Sheridan, P. & Lu, W. D. Data clustering using memristor networks. Sci. Rep. 5, 10492 (2015).

22. Prezioso, M. et al. Training and operation of an integrated neuromorphic network based on metal-oxide memristors. Nature 521, 6164 (2015).

23. Tsymbal, E. Y. Tunneling across a ferroelectric. Science 313, 181183 (2006).24. Garcia, V. & Bibes, M. Ferroelectric tunnel junctions for information storage and processing. Nat. Commun. 5, 4289 (2014).

25. Chanthbouala, A. et al. A ferroelectric memristor. Nat. Mater. 11, 860864 (2012).

26. Yamada, H. et al. Giant electroresistance of super-tetragonal BiFeO3-based ferroelectric tunnel junctions. ACS Nano 7, 53855390 (2013).

27. Boyn, S. et al. High-performance ferroelectric memory based on fully patterned tunnel junctions. Appl. Phys. Lett. 104, 52909 (2014).

28. Linares-Barranco, B. & Serrano-Gotarredona, T. Memristance can explain spike-time-dependent-plasticity in neural synapses. Preprint at http://precedings.nature.com/documents/3010/version/1

Web End =http:// http://precedings.nature.com/documents/3010/version/1

Web End =precedings.nature.com/documents/3010/version/1 (2009).

29. Gruverman, A. et al. Direct studies of domain switching dynamics in thin lm ferroelectric capacitors. Appl. Phys. Lett. 87, 82902 (2005).

6 NATURE COMMUNICATIONS | 8:14736 | DOI: 10.1038/ncomms14736 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications

NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14736 ARTICLE

30. Bhattacharjee, S., Rahmedov, D., Wang, D.,niguez, J. & Bellaiche, L. Ultrafast switching of the electric polarization and magnetic chirality in BiFeO3 by an electric eld. Phys. Rev. Lett. 112, 147601 (2014).

31. Tagantsev, A., Stolichnov, I., Setter, N., Cross, J. & Tsukada, M. Non-Kolmogorov-Avrami switching kinetics in ferroelectric thin lms. Phys. Rev. B 66, 214109 (2002).

32. Jo, J. Y. et al. Domain switching kinetics in disordered ferroelectric thin lms. Phys. Rev. Lett. 99, 267602 (2007).

33. Strukov, D. B. & Williams, R. S. Four-dimensional address topology for circuits with stacked multilayer crossbar arrays. Proc. Natl Acad. Sci. USA 106, 2015520158 (2009).

34. LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436444 (2015).35. Marinova, M. et al. Depth proling charge accumulation from a ferroelectric into a doped mott insulator. Nano Lett. 15, 25332541 (2015).

36. Wang, D., Weerasinghe, J. & Bellaiche, L. Atomistic molecular dynamic simulations of multiferroics. Phys. Rev. Lett. 109, 67203 (2012).

37. Kornev, I. A., Lisenkov, S., Haumont, R., Dkhil, B. & Bellaiche, L. Finite-temperature properties of multiferroic BiFeO3. Phys. Rev. Lett. 99, 227602 (2007).

38. Albrecht, D. et al. Ferromagnetism in multiferroic BiFeO3 lms: a rst-principles-based study. Phys. Rev. B 81, 140401 (R) (2010).

Acknowledgements

We thank Cyrile Deranlot and Eric Jacquet for assistance with the sample fabrication and D.J. Kim for fruitful discussions. Financial support from the European Research Council (ERC Advanced Grant No. 267579) and the French Agence Nationale de la Recherche (ANR) through projects FERROMON and MIRA are acknowledged. This publication has received funding from the European Unions Horizon 2020 research innovation programme under grant agreement 732642 (ULPEC project). This work is supported by a public grant overseen by the French National Research Agency (ANR) as part of the Investissements dAvenir program (Labex NanoSaclay, reference: ANR-10-LABX-0035). B.X. and L.B. acknowledge the nancial support from the Department of Energy, Ofce of Basic Energy Sciences, under contract ER-46612, and DARPA grant HR0011-15-2-0038 (MATRIX program), respectively.

Author contributions

J.G., S.S. and V.G. designed and organized the experiments and calculations. S.G., C.C., K.G. and S.X. elaborated the thin lms and nanofabricated the samples. S.F. and V.G. performed the piezoresponse force microscopy experiments. S.B., J.G., G.L., S.F., J.T., S.S. and V.G. performed the electrical transport measurements. B.X. performed the atomic-scale molecular dynamics simulations, which were then discussed with L.B. S.B., N.L. and J.G. performed the spiking neural network simulations. All the authors contributed to the interpretation of the results. S.B., J.G. and V.G. wrote the manuscript with input from all authors.

Additional information

Supplementary Information accompanies this paper at http://www.nature.com/naturecommunications

Web End =http://www.nature.com/ http://www.nature.com/naturecommunications

Web End =naturecommunications

Competing interests: The authors declare no competing nancial interests.

Reprints and permission information is available online at http://npg.nature.com/reprintsandpermissions/

Web End =http://npg.nature.com/ http://npg.nature.com/reprintsandpermissions/

Web End =reprintsandpermissions/

How to cite this article: Boyn, S. et al. Learning through ferroelectric domain dynamics in solid-state synapses. Nat. Commun. 8, 14736 doi: 10.1038/ncomms14736 (2017).

Publishers note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional afliations.

This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the articles Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/

Web End =http://creativecommons.org/licenses/by/4.0/

r The Author(s) 2017

NATURE COMMUNICATIONS | 8:14736 | DOI: 10.1038/ncomms14736 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 7