University of Groningen
A non-volatile organic electrochemical device as a low-voltage artificial synapse for
neuromorphic computing
van de Burgt, Yoeri; Lubberman, Ewout; Fuller, Elliot J.; Keene, Scott T.; Faria, Gregorio C.;
Agarwal, Sapan; Marinella, Matthew J.; Talin, A. Alec; Salleo, Alberto
Published in:
Nature Materials
DOI:
10.1038/NMAT4856
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from
it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date:
2017
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
van de Burgt, Y., Lubberman, E., Fuller, E. J., Keene, S. T., Faria, G. C., Agarwal, S., Marinella, M. J.,
Talin, A. A., & Salleo, A. (2017). A non-volatile organic electrochemical device as a low-voltage artificial
synapse for neuromorphic computing.
Nature Materials
,
16
(4), 414-418. https://doi.org/10.1038/NMAT4856
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the
author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
The publication may also be distributed here under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license.
More information can be found on the University of Groningen website: https://www.rug.nl/library/open-access/self-archiving-pure/taverne-
amendment.
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the
number of authors shown on this cover page is limited to 10 maximum.
LETTERS
PUBLISHED ONLINE: 20 FEBRUARY 2017 |
DOI: 10.1038/NMAT4856
A non-volatile organic electrochemical device as a
low-voltage artificial synapse for neuromorphic
computing
Yoeri van de Burgt
1†‡
, Ewout Lubberman
1,2‡
, Elliot J. Fuller
3
, Scott T. Keene
1
, Grégorio C. Faria
1,4
,
Sapan Agarwal
3
, Matthew J. Marinella
5
, A. Alec Talin
3
*
and Alberto Salleo
1
*
The brain is capable of massively parallel information pro-
cessing while consuming only ∼1–100 fJ per synaptic event
1,2
.
Inspired by the efficiency of the brain, CMOS-based neural
architectures
3
and memristors
4,5
are being developed for pat-
tern recognition and machine learning. However, the volatil-
ity, design complexity and high supply voltages for CMOS
architectures, and the stochastic and energy-costly switching
of memristors complicate the path to achieve the intercon-
nectivity, information density, and energy efficiency of the
brain using either approach. Here we describe an electrochem-
ical neuromorphic organic device (ENODe) operating with a
fundamentally different mechanism from existing memristors.
ENODe switches at low voltage and energy (<10 pJ for 10
3
µm
2
devices), displays >500 distinct, non-volatile conductance
states within a ∼1 V range, and achieves high classification ac-
curacy when implemented in neural network simulations. Plas-
tic ENODes are also fabricated on flexible substrates enabling
the integration of neuromorphic functionality in stretchable
electronic systems
6,7
. Mechanical flexibility makes ENODes
compatible with three-dimensional architectures, opening a
path towards extreme interconnectivity comparable to the
human brain.
Two-terminal memristors based on filament-forming metal
oxides (FFMOs) or phase change memory (PCM) materials have
recently been demonstrated to function as non-volatile memory
that can emulate neuronal and synaptic functions such as long-term
potentiation (LTP), short-term potentiation (STP), and spike timing
dependent plasticity (STDP)
4,5
. Crossbar architectures based on
these devices have been projected to reduce energy costs for neural
algorithms by six orders of magnitude, and recently performed
image recognition and data classific ation when utilized as highly
parallel neuromorphic processing units
8,9
. However, despite recent
progress in the fabrication of device arrays, to date no architecture
has been shown to operate with the projected energy efficiency
while maintaining high accuracy. A major impediment still exists
at the device level; specifically, a resistive memory device has
not yet been demonstrated with adequate electrical characteristics
to fully realize the efficiency and performance gains of a neural
architecture. State-of-the-art memristors suffer from excessive
write noise
10
, write nonlinearities
8
and high write voltages and
currents
11
. Reducing the noise and lowering the switching voltage
significantly below 0.3 V (∼10 kT) in a two-terminal device without
compromising long-term data retention has proven difficu lt
12
.
These limitations reduce the accuracy and scalability of FFMO and
PCM memristors and pose challenges for these devices to approach
the energy efficiency of the brain
8
.
Recognizing that different switching mechanisms may be bene-
ficial, organic memristive devices have b een recently proposed
13–15
.
Besides low-cost manufacturing and flexibility inherent to soft ma-
terials, organic devices could als o benefit from low-power consump-
tion, added functionality, and biocompatibility. They could act as
biometric sensors and direct interfaces with the brain
16,17
, opening
up the tantalizing opportunity to build advanced neural prosthe-
ses comprising integrated brain–machine interfaces that combine
neural sensing with training
18
. However, the operation of these
organic memristors relies either on the slow kinetics of ion diffusion
through a polymer to retain their states or on charge storage in metal
nanoparticles, which inherently limits performance and stability.
In contrast, the operation of ENODe is based on the non-volatile
control of the conductivity of an organic mixed ionic/elect ronic
conductor as depicted in Fig. 1. ENODe is essentially similar
to a concentration battery. During the ‘read’ operation, the
cell is disconnected and the electronic charge of t he electrodes
remains unaltered by virtue of an ion conducting/electron blocking
electrolyte. The charge in the electrodes is manipulated during the
‘write’ operation. Hence, ENODe is a type of non-volatile redox
cell (NVRC) in which the state of charge determines the electronic
conductivity
19
. The main advantage of NVRCs is that the barrier
for state retention is decoupled from the barrier for changing states,
allowing for the extremely low switching voltages while maintaining
non-volatility (Fig. 1c).
To demonstrate this concept, we use a poly(3,4-ethylene-
dioxythiophene):polystyrene sulfonate (PEDOT:PSS) film part ially
reduced with poly(ethylenimine) (PEI) (see Methods). The three-
terminal de vice architecture comprises the postsynaptic electrode,
a PEI/PEDOT:PSS film, interfaced with a PEDOT:PSS presynaptic
electrode via an electrolyte (Fig. 1a). Upon applying a positive
presynaptic potential V
pre
to the PEDOT:PSS electrode, cations
flow from the presynaptic electrode into the postsynaptic electrode
through the electrolyte, resulting in protonation of the PEI, while
electrons flow through the external circuit. This causes holes
to be removed from the PEDOT backbone in the postsynaptic
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
1
Department of Materials Science and Engineering, Stanford University, Stanford, California 94305, USA.
2
Zernike Institute for Advanced Materials,
University of Groningen, 9747AG Gronigen, The Netherlands.
3
Sandia National Laboratories, Livermore, California 94551, USA.
4
Instituto de Física de São
Carlos, Universidade de São Paulo, 13566-590 São Carlos, SP, Brasil.
5
Sandia National Laboratories, Albuquerque, New Mexico 87123, USA.
†
Present
address: Microsystems and Institute for Complex Molecular Systems, Eindhoven University of Technology, 5612AJ Eindhoven, The Netherlands.
‡
These
authors contributed equally to this work.
*
e-mail:
aatalin@sandia.gov; asalleo@stanford.edu
414 NATURE MATERIALS | VOL 16 | APRIL 2017 | www.nature.com/naturematerials
NATURE MATERIALS DOI: 10.1038/NMAT4856
LETTERS
PEDOT
PEI
PSS
Oxidation
Reduction
V
pre
Bias V
pre
E
Open
Closed/write
Closed
OCP (V)
State of charge State of charge
V
post
DrainSource
I
post
SO
3
H
S
SO
3
H
O
OO
S
S
OO
OO
S
S
OO
OO
H
2
N
H
N
SO
3
H
S
SO
3
H
O
OO
S
S
OO
OO
S
S
OO
O
O
H
+
+
+
−
−
G (μS)
V
pre
a
PEDOT:PSS
PEDOT:PSS/PEI
H
+
H
+
e
−
Open/read
E
e
−
eV
b
1 2
1
2
900
800
700
600
500
400
12345678910
Pulse #
cd
b
Figure 1 | Structure and electronic states of an organic neuromorphic device. a, Sketch of the device structure. Pre- and postsynaptic layers are separated
by an electrolyte layer transporting ions/protons (red spheres). b, A positive V
pre
drives protons into the postsynaptic electrode, which results in the
compensation of some PSS
−
by the protonated PEI. This reaction causes the reduction of PEDOT in the same electrode due to charge neutrality, which
eliminates a polaron (in red) and decreases the polymer conductivity. The reaction is reversed upon applying a negative V
pre
. c, Schematic explaining the
decoupling of the read and write operations. NVRC ensures a very high eV
b
barrier between the two oxidation states of PEDOT ‘1’ and ‘2’ (corresponding to
two conductance states of the postsynaptic electrode) during an open read operation and a very low barrier during a closed write operation. The open
circuit potential (OCP), depicted in dashed lines, is dependent on the oxidation state of PEDOT and can be overcome by the bias. d, Conductance G of the
postsynaptic electrode, showing reproducible non-volatile switching between five discrete states.
electrode, thereby reducing its electronic conductivity while
ensuring electroneutrality in the electrode (Fig. 1b). The reaction
is reversed upon applying a negative V
pre
. While enabling current
continuity by ion transport, the electrolyte also acts as a barrier for
electronic charge transport, maintaining the electrode conductance
state after the presynaptic potential is applied. PEI stabilizes the
neutral form of the PEDOT in the PEDOT:PSS/PEI electrode,
ensuring that the oxidation state of the postsynaptic electrode
is retained
20
. The conductance states are monitored using a
postsynaptic potential V
post
. As such, the conductance of the
PEDOT:PSS/PEI channel represents the synaptic weight of the
connection between two neurons
21
, an essential property of an
artificial synapse.
We show that ENODe exhibits some of the synaptic functions
that are the building blocks of neuromorphic computing. To
demonstrate the extremely high density of non-volatile states
available for computation, a series of 500 pulses are applied
(see Methods and Supplementary Information), resulting in 500
distinct conductance states (Fig. 2a). In addition to driving it
with V
pre
, ENODe can b e operated by injecting a presynaptic
current pulse (Fig. 2b) exhibiting a nearly perfect linear behaviour.
We cycled ENODe between two distinct states over 300 times
using 10 mV potentiation and depotentiation pulses, demonstrating
extremely low noise (Supplementary Fig. 1) (<1%), which enables
the definition of a large number of states in a small voltage range.
The postsynaptic state is programmed by varying the amplitude
or the duration of the presynaptic pulse. The conductance change,
1G, is a linear function of presynaptic pulse amplitude and
duration (Fig. 2d,e), down to approximately millisecond timescales
(see inset).
Below 6 ms (>166 Hz) the potentiation is only short term
(Supplementary Fig. 3). This timescale is consistent with a
diffusion time constant, τ ∼ L
2
/D of ∼10 ms, estimated using a
previously me asured charge carrier diffusivity in PEDOT:PSS of
∼10
−8
cm
2
s
−1
and an electrode thickness of ∼100 nm (ref. 22).
Reducing the channel thickness will reduce the diffusion distance
and improve the time response. As sub-threshold potentiation
in neurons is ass ociated with STP and paired pulse facilitation
(PPF), this functionality is also established in ENODe (Fig. 2c
and Supplementary Fig. 2). Interestingly, the PPF demonstrated
in Fig. 2c exhibits two characteristic timescales, τ
1
= 14 ms and
τ
2
= 240 ms, approximately equal to those measured in biological
synapses
23
. Addit ional bio-inspired functionality such as STDP can
be achieved using overlapping pulse design (see Supplementary
Information). Although STP is capacitive in nature, applying many
short pulses results in LTP (Supplementary Fig. 3), a behaviour
emulating short-term to long-term potentiation found in nature
24
.
Size and geometry not only dictate operating speed, but also
define switching energy. To highlight the path towards ultralow-
energy switching of ENODe, power dissipation was measured in
devices with areas spanning five orders of magnitude (see Fig. 2f).
The power dissipated is determined by P = I × V , and the energy
is calculated by integration over the pulse width (Supplementary
Fig. 4a). The switching energy of our smallest device was mea-
sured to be ∼10 pJ, which is comparable to state-of-the-art PCMs
that are over three orders of magnitude smaller, demonstrating the
NATURE MATERIALS | VOL 16 | APRIL 2017 | www.nature.com/naturematerials
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
415
LETTERS
NATURE MATERIALS
DOI: 10.1038/NMAT4856
0 250 500 750 1,000 1,250 1,500 1,750 2,000
0.0
0.1
0.2
0.3
∆t (ms)
∆t
a
0102030405060708090
2,500
2,000
1,500
1,000
Pulse #
0 2,500 5,000 7,500 10,000 12,500 15,000
800
600
4,000 4,500 5,000 5,500
800
750
700
650
600
550
−1
0
1
G (µS)
G (µS)
Conductance, G (µS)
∆G (µS)
∆G (µS)
∆G (µS)
∆G
−80
−75
−70
−65
−60
−55
Postsynaptic current (µA)
V
pre
(mV)
Pulse #
V
post
= 100 mV
02468
0
40
80
120
0
100 200
0
3
6
9
010203040
50
0
150
300
450
Presynaptic pulse amplitude (mV)
Presynaptic pulse length (s)
5 mV presynaptic pulse
2 s presynaptic pulse
Pulse length (ms)
10 mV
10
−2
10
−3
10
−2
10
−1
10
0
10
1
10
2
10
−1
10
0
10
1
10
2
Energy (nJ)
Area (mm
2
)
Time
Pre-
synaptic
pulse
Post-
synaptic
current
PPF = C
1
e
t/τ
1 + C
2
e
t/τ
2
τ
1
= 14 ms
τ
2
= 240 ms
b
d
e
f
c
Figure 2 | Neuromorphic behaviour. a, Long-term potentiation and depression displaying 500 discrete states over the operating range when the device is
controlled using voltage pulses. The inset is a zoom-in showing the individual states. b, Long-term potentiation and depression under current control.
c, Short-term potentiation and paired pulse facilitation. The amount by which the synaptic weight is temporarily modified depends on the time interval
between two short pulses. An exponential fit is applied to obtain two characteristic timescales. The inset is a schematic of how such biasing is typically
realized. d,e, Change in postsynaptic conductance as a function of presynaptic pulse amplitude (d) and duration (e). The inset in e shows the relationship
for shorter timescales and V
pre
= 10 mV. f, Switching energy measured as a function of device area. Linear fits are applied in d to f.
partic ularly low switching energy density of ENODe
25
. Since current
scales with area whereas the voltage, determined by the electro-
chemical overpotential at the polymer/electrolyte interface, remains
approximately constant, the switching energy is proportional to the
electrode area, with a slope of 390 ± 10 pJ mm
−2
(Fig. 2f). Thus,
we project an energy cost of 35 aJ for switching a 0.3 × 0.3 µm
device, which can be fabricated by photolithography
26
. Downscal-
ing of ENODe will also require that more resistive PEDOT:PSS
formulations be used. We demonstrated using such formulations
to fabricate devices with conductances ranging over three orders
of magnitude (see Supplementary Information and Supplementary
Figs 5,6). The energy advantage of ENODe is further enhanced by
the low switching voltage (∼0.5 mV), which greatly reduces the
interconnect capacitive loss in arrays and is ∼×10
3
lower than the
‘write’ voltage for a typical memristor.
Taking advantage of processing techniques developed for
commodity p olymers, we fabricated an all solid-state plastic
device. Nafion was used as the electrolyte, laminated between two
flexible PEDOT:PSS films coated on polyethylene terephthalate
(PET) sheets and permeated with PEI (Fig. 3). This all-plastic
device proves the potenti al for low-cost fabrication of flexible
ENODe arrays, which would enable the integration of on-bo ard
neuromorphic computing and learning in implantable prosthetics,
neural electrode arrays or any other flexible large-area electronic
system
17
. Furthermore, bending and folding of arrays may enable
three-dimensional densely connected neuromorphic devices.
As a first simple demonstration of functionality, we integrated
ENODe in a circuit that emulates Pavlovian learning
27
(Fig. 4a and
Supplementary Information). The output neuron N3 (salivation) is
triggered by the input neuron N1 (sight of food, panel 1) but initially
not by neuron N2 (bell ringing, panel 2). The synaptic weight
of ENODe (S2) is modified during learning, thereby permanently
associating N2 to N1 (panel 3). The learning process resulted
in a response at N3 (salivation), to the input N2 (bell ringing),
successfully demonstrating associative memory of our artificial
synapse (panel 4).
Further, to fully illustrate the power of the low noise and linearly
programmable conductance states of ENODe, we simu lated a
neural network based upon its exp erimentally measured properties
(see Supplementary Information). We simulated a three-layer
network for training with back-propagation of three data sets:
an 8 × 8 pixel image version of handwritten digits
28
; MNIST, a
28 × 28 pixel version of handwritten digits
29
; and a Sandia file
classification data set
30
. Back-propagation is a well-studied method
that provides benchmarking with the data sets we used
8
. The
numerical weights in the network layer were mapped directly
onto the experimental device conductance states (Fig. 2a) that are
extracted from ∼15,000 experimentally measured states. Training
a neural network using ENODe gives an accuracy between 93%
and 97%, and is always within 2% of the ideal floating-point-
based neural network performance, which is the theoretical limit
for this algorithm (Fig. 4d–f). Using a similar algorithm on PCM
devices previously yielded far lower classification accuracies
8
. The
key to this exceptional performance is the linearity and low noise
of ENODe (Fig . 2d,e), allowing ext remely efficient analog tuning
31
.
In contrast, t he physics of switching PCMs and FFMOs imposes
416
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
NATURE MATERIALS | VOL 16 | APRIL 2017 | www.nature.com/naturematerials
NATURE MATERIALS DOI: 10.1038/NMAT4856
LETTERS
2,000 2,500 3,000 3,500 4,000 4,500
1,900
1,850
1,800
1,750
Time (s)
Conductance (μS)
a
Postsynaptic electrode
Presynaptic electrode
1,860
1,850
1,840
ΔG = 3 µS
PET substrate
PEDOT:PSS
presynaptic electrode
Nafion electrolyte
PEDOT:PSS/PEI
postsynaptic electrode
b
c
Figure 3 | Flexible all solid-state neuromorphic device. a, Schematic of the device. b, Photograph of the device while being flexed. c, 125 potentiation and
depotentiation states obtained with 0.5 mV pulses. The inset shows the conductance difference between contiguous states.
0
10 20 30 40 50 60
0
8
10
0
100
200
0
8
16
0
100
200
Time (s)
Postsynaptic current (µA)
Probing ‘Bell ringing’
Associating
‘Bell ringing’ with
’Sight of food’
Probing ‘Bell ringing’
Probing ‘Sight of food’
S
1
S
2
Sight of food
Salivation
−1
0
010203040
Accuracy
Training Epoch
File types Small digits Large digits
Exp. derived
Ideal numeric
Exp. derived
Ideal numeric
Exp. derived
Ideal numeric
90
99
0
010203040
Accuracy
Training Epoch
90
99
0
010203040
Accuracy
Training Epoch
90
99
−2
−1
0
1
600 700 800
0
0.0
0.2
0.4
0.6
0.8
1.0
1
2
Bell ringing
N
1
N
2
N
3
a
∆G (µS)
∆G (µS)
Conductance (µS)
600 700 800
Conductance (µS)
CDF
0.0
0.2
0.4
0.6
0.8
1.0
CDF
b
def
c
Figure 4 | Learning circuit and image recognition simulations. a, Schematic and results of the Pavlovian learning circuit. Conditioning and permanent
association are shown in the third and fourth panel. b,c, Heat map representation of the 1G versus G switching statistics of ENODe during potentiation (b)
and depotentiation (c). The heat maps encompass data from 15,000 measurements and their colour represents the cumulative distribution function (CDF)
at each conductance state. CDF is the probability that 1G is less or equal to the 1G plotted. d–f, Backpropagation training results using a Sandia file
classification data set (d), an 8 × 8 pixel handwritten digit image (e) and a 28 × 28 pixel handwritten digit image (f).
inherently nonlinear device characteristics as it relies on random
nucleation events (PCM) or on modulating a tunnel barrier over a
narrow region (FFMO)
8,10
.
Analog tuning and extremely low switching voltages are
consistent with the inherently fast and low-energy process of
ion transport into a swollen polymer, requiring only a small
electrochemical overpotential
32
. We compared the composition of
the pre- and postsynaptic electrodes before and after operation
in liquid electrolyte (KCl) using X-ray photoelectron spectroscopy
(XPS) and propose that cat ions from the supporting electrolyte
(K
+
) are mobile in the presynaptic electrode whereas protons
(H
+
) are the mobile species in the postsynaptic electrode (see
Methods and Supplementary Information). Thus, upon applying
a positive V
pre
, K
+
cations are emitted from the presynaptic
electrode into the electrolyte. As a consequence of this increase in
positive charge concentration in the electrolyte, protons penetrate
the PEDOT:PSS/PEI postsynaptic electrode and protonate amine
groups in the PEI, which due to charge neutrality reduces
the concentration of PEDOT
+
, de creasing its conductivity. The
opposite process occurs upon applying a negative V
pre
.
The mechanism, which differs fundamentally from that of or-
ganic neuromorphic devices reported to date, explains the origin of
the non-volatile nature of the conductance states of our device
13,14
.
Electroneutrality in the device imposes that any charging caused
by biasing the presynaptic electrode is balanced by an equal dop-
ing/dedoping in the postsynaptic electrode, resulting in a continu-
ous analog tuning of the ENODe conductivity. Upon programming
the device to a specific state, the presynaptic PEDOT:PSS electrode
and the postsynaptic PEDOT:PSS/PEI film are at different potentials
caused by the different PEDOT
+
concentration. Once programmed,
the electrodes are disconnected from sources of charge and have no
direct elec trical connection to each other either since the electrolyte
NATURE MATERIALS | VOL 16 | APRIL 2017 | www.nature.com/naturematerials
© 2017 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.
417