Stochastic networks of memristive devices were fabricated using a sponge as a skeleton material. Cyclic voltage-current characteristics, measured on the network, revealed properties, similar to the organic memristive device with deterministic architecture. Application of the external training resulted in the adaptation of the network electrical properties. The system revealed an improved stability with respect to the networks, composed from polymer fibers.

Memristive devices1–3 are currently considered as perspective elements for the realization of electronic circuits mimicking some features of the nervous system elements of living beings.4 It has been reported several examples when memristive devices were used as electronic analogs of neurons5,6 and synapses.7,8 In particular, using the organic memristive device9 it was reconstructed an electronic circuit, mimicking the architecture and properties of a part of the nervous system of the pond snail Lymnea stagnalis, responsible for the learning of this animal during feeding.10 In this case Hebbian (synaptic) type of learning11 was demonstrated in artificially fabricated electronic circuits. The realized system was rather simple as only 2 synapses were used for the circuit fabrication. Circuits, mimicking the behavior of the Pavlov’s dog were also reported using memristive devices.12 However, in this last case only the main feature of the animal was reproduced, disregarding the system architecture, because it is very complicated and precise data are not available.

Realization of more complicated neuromorphic networks will require a significant increase of the number of memristive devices. Current electronic technologies allow to minimize elements sizes and to increase an integration degree. However, these technologies have some limitations regarding the fabrication of 3D structures – an essential feature of the brain organization.

There are several realizations of the organic memristive systems.13,14 In,13 authors report synapse-like behavior, connected to the barrier formation between nanoparticles and organic material. The barrier formation approach was used also in the case of the stochastic polymeric networks,15 but for other reasons – distribution of threshold elements in the network.

In this respect, organic materials can be very useful as they allow formation of self-assembled 3D systems: a large variety of reported organic and composite structures, realized by this method, indicates the possibility of the realization of practically any type of the molecular architectures. The approach can be applied also for the realization of neuromorphic networks. In this case, the developed method must result in the fabrication of stochastically organized system with the distributed contacts of at least two different materials. Each contact area must have characteristics similar to those of the deterministically assembled organic memristive devices, allowing, therefore, to vary the strength (conductivity) of connections between different input-output electrodes pairs by the application of the appropriate external training algorithms.

Two types of stochastic organic systems were reported: one was based on polymer fiber structures16 and the other one was constructed using phase separation of specially synthesized block-copolymers.15 Both systems have demonstrated adaptation capabilities. However, in the case of the fiber structure the stability of the electrical properties was found to be very low. Such behavior was attributed to the free-standing nature of the fibers in the fabricated systems; their deformation and even breaking took place due to the local heating during the current passing. In the second case, the properties were much more stable. Moreover, several similarities to the living beings learning were revealed. However, this approach requires a synthesis of new materials, what makes more complicated the network fabrication procedure. The present paper is dedicated to the development of more simple method of the preparation of the stochastic systems, allowing adaptations and learning as the result of the training procedure, with better stability of the electrical properties.

Comparing the two mentioned examples, we need to combine somehow their main advantages: simplicity of the fiber approach and stability of the self-assembling structures. Therefore, it is better to avoid the synthesis of complicated molecules and to realize desired systems from commercially available compounds. In order to exclude the degradation of the conducting pathways properties and to guarantee, therefore, an increased stability of the network, we have followed the way of nature: stabilization of soft matter (polymers) organization and properties requires the use of skeletons – rigid and stable supports. In order to have the possibility to realize stochastic 3D systems, these supports must have a porous developed structure, where the required materials can penetrate. In this work we have used sponge for these reasons.

Other considerations of the stochastic networks are relater to the statistical growth of filaments17 and combination of CMOS-memristor systems.18 

It is to note that such systems are expected to reveal improved noise resistance.19,20

Organic memristive devices (our basic element for the adaptive networks) is formed by a heterojunction of conducting polymer and solid electrolyte, in particular, polyaniline (PANI) and polyethylene oxide (PEO) doped with lithium salts.9 Thus, the formed supported network must contain these materials distributed in a stochastic manner with a statistical probability of forming the desirable contacts of two mentioned materials.

Scheme of the sample is shown in Fig. 1(a).

FIG. 1.

(a) Scheme of the sample used for the study of the training of the stochastic network, assembled on porous support; (b) Photo of the sponge support, covered by PANI (size 15 mm × 10 mm × 5 mm).

FIG. 1.

(a) Scheme of the sample used for the study of the training of the stochastic network, assembled on porous support; (b) Photo of the sponge support, covered by PANI (size 15 mm × 10 mm × 5 mm).

Close modal

The main element (functionalized sponge) was prepared in the following way. Natural cellulose sponge was used as a skeleton in this work. The sponge had the following sizes: length – 15 mm; width – 8 mm; height – 5 mm. The medium cavity size was of in the range between 0.1 mm - 0.5 mm. The optical image of the sponge is shown in Fig. 1(b). Initially, in was immersed into 0.1 M HCl for 10 hours. After it, the sponge was dried by its placing on the filter paper and immersed into the solution of PANI for 30 minutes. The PANI solution was prepared according to Ref. 21. Emeraldine base of PANI was solved in dimethylacedamide at a concentration of 20 mg/ml. Stirring of the solution was done during 14 hours followed by the sonication during 10 hours. The resultant solution was filtered. One part of the filtered PANI solution, prepared as described above, was added to nine parts of water with pH 3.2, adjusted by HCl adding. After dissolving, the pH of the solution was decreased till 2.5 – 2.6 by adding drops of 1M HCl. Then, the sample was dried by placing it on the filter paper. PEO (8000000 Da, Sigma) was dissolved in water (Milli-Q) solution of 0.1 M LiClO4 and 0.1 M HCl at a concentration of 20 mg/ml. 0.1 ml of this solution was injected in the center of the PANI covered sponge sample by syringe. After it, the sample was placed into the exiccator and vacummated during 30 minutes by the mechanical pump. The formation and collapse of the PEO bubbles was clearly visible on the surface of the sponge support. After it, three gold and one silver electrodes were mechanically inserted into the sample, as it is shown in Fig. 1(a).

One of gold wires (it can be called source) is connected to the ground potential level. The other two gold wires are drains and each of them is connected to the independent power supply through amperometers.

As the conductivity switching in organic memristive device is controlled by the redox state of PANI,22 the structure must be supplied by the reference electrode. This electrode (R - silver wire) is also inserted into the stochastic matrix and connected to the ground potential. Another ampermeter was inserted into the circuit of the reference electrode in order to measure the ionic current in the system.

Measurements were done in the closed chamber with residual HCl vapors (a backer with 0.1 ml of concentrated HCl was placed into the chamber).

Cyclic voltage current characteristics of the conducting branches of the network were measured in the following way. Voltage sweep and drain current measurements were performed with 236 source measure units (Keithley). The gate current is registered with a Keithley 6514 system electrometer. The Keithley units were connected to a PC and operated via MATLAB scripts, which allowed full automation of the measuring procedure.

Before experiments on the network training we have measured cyclic voltage current characteristics in each possible signal pathway branch of the network (S-D1 (application of V1 and current measurement with A1) and S-D2 (application of V2 and current measurement with A2) respectively). In parallel, cyclic characteristics for the ionic current in the chain of the reference electrode were also measured (using AR). Measurements were started from 0 V, increasing till the maximum positive voltage of +1.2 V with the step of 0.1 V. After it, the voltage was decreased till -1.2 V and increased again till 0 V with the same step of 0.1 V. Values of the current were readout after 60 s delay after application of each voltage value. These characteristics are shown in Fig. 2 and their behavior can be effectively described by the model of the organic memristive device functioning.23 

FIG. 2.

Cyclic voltage-current characteristics, measured between each S-D terminals in the circuit shown in Fig. 1 for total (a) and ionic (b) currents. Time delay between the voltage application and the current value readout was 60 s. Arrows indicate the direction of the voltage changing.

FIG. 2.

Cyclic voltage-current characteristics, measured between each S-D terminals in the circuit shown in Fig. 1 for total (a) and ionic (b) currents. Time delay between the voltage application and the current value readout was 60 s. Arrows indicate the direction of the voltage changing.

Close modal

The characteristics shown in Fig. 2 demonstrate the realization of the suitable PANI-PEO mutually arranged junctions. The presence of both important features, namely, hysteresis and rectification, is clearly visible in the figure. The other essential feature of the presented characteristics is connected to the fact that the values of the ionic current is about one order of magnitude lower with respect to the total current value. It means that the composition of the solid electrolyte in the network was optimal – the concentration of ions, whose motion is necessary for the conductivity switching,24 is enough for the effective carrying out of the redox reactions, but the presence of these ions does not contribute significantly to the value of the total device current (stochastic nature of the network does not allow to have more pronounced difference between the values of these current; it can reach about three orders of magnitude in devices with a deterministic architecture).4 The cyclic characteristics for the differential current, representing the electronic compound of the total current of the single branch of the stochastic network, is shown in the Fig. 3.

FIG. 3.

Cyclic voltage current characteristics (difference between the total and ionic currents, shown in Fig. 2) as a function of the applied voltage. Arrows indicate the direction of the voltage changing.

FIG. 3.

Cyclic voltage current characteristics (difference between the total and ionic currents, shown in Fig. 2) as a function of the applied voltage. Arrows indicate the direction of the voltage changing.

Close modal

Comparing the data presented in Figs. 2 and 3 with those, measured on networks, fabricated with fiber polymeric structures,16 we can see a significant difference – characteristics shown in Figs. 2 and 3 reveal not only rectification, but also a pronounced hysteresis. It is a first indication of the fact that the system is much more stable. In fact, weak hysteresis in the case of fiber polymer structures can be connected to the banding or breaking of the continuity of fibers. Thus, the characteristics observed in the case of fibrillar structures resulted from the action of two factors: application of external electrical stimuli and mechanical deformation of the network due to the local heating of the structure elements.

Once making a proof that each branch of the system can work as a single organic memrisrive device, the experiments on the system adaptations and learning were carried out. Recently, learning in the memristor-based networks is under a wide discussion.25–31 

We assumed that the training must result in the reinforcement of the conductivity in one branch (S-D1) and to inhibit it in the other one (S-D2)

Similarly to the previously reported systems,15,16 three different signals were applied to the network: two types of training stimuli and testing signal. Testing signal value was chosen in such a way that it does not varies significantly the conductivity state of the system (+0.3 V). Training stimuli, instead, must vary these states. Therefore, the S electrode was maintained always at a ground potential value, while the potential of +1.0 V was applied between S-D1 and the potential – 0.2 V was applied between S-D2 simultaneously for 60 min.

The conductivity between these pathways was measured before and after the described training procedure and the results are summarized in the Table I.

TABLE I.

Results of the stochastic system learning during the application of the external stimuli.

S-D1 S-D2
Current value for the applied test voltage before training (nA)  50  55 
Current value for the applied test voltage after training (nA)  200  21 
S-D1 S-D2
Current value for the applied test voltage before training (nA)  50  55 
Current value for the applied test voltage after training (nA)  200  21 

Results, shown in Table I, clearly demonstrate learning of the system and adaptation of its electrical properties.

It seems important that the conductivity ratio between the reinforced and inhibited branches in this case is mush higher than in the case of systems of free-standing fibers. If we make a comparison with systems, based on phase-separating self-assembling structure,15 the ratio here is comparable with so called “adult” learning case, and about one order of magnitude less than in the case of so called “baby” learning (imprinting). However, considering that the presented approach is very simple, we can suggest that the further works in this direction, connected to the better choice of the skeleton template, optimization of the deposition techniques and variation of the device architecture, will allow to fabricate networks with high performance of characteristics. These characteristics correspond well to the distributed mnemotrixes in the Valentino Braitenberg’s mental experiment, explaining learning.32 Our stochastically realized connections allow reinforcement and inhibition, resulting in the structuring of properties of the initially stochastic networks.

The work is performed according to the Russian Government Program of Competitive Growth of Kazan Federal University. The work was partially supported by the MaDEleNA project financed by the Provincia Autonoma di Trento, call “Grandi Progetti 2012.”

1.
L. O.
Chua
,
IEEE Trans. Circuit Theory
18
,
507
(
1971
).
2.
D. B.
Strukov
,
G. S.
Snider
,
D. R.
Stewart
, and
R. S.
Williams
,
Nature
453
,
80
(
2008
).
3.
Y. V.
Pershin
and
M.
Di Ventra
,
Adv. Phys.
60
,
145
(
2011
).
4.
V.
Erokhin
and
M. P.
Fontana
,
J. Computational Theor. Nanosci.
8
,
313
(
2011
).
5.
M. D.
Pickett
,
G.
Medeiros-Ribeiro
, and
R. S.
Williams
,
Nature Materials
12
,
114
(
2013
).
6.
M.
Poggio
and
L. P.
Lee
,
Biophys. J.
106
,
795a
(
2014
).
7.
V.
Erokhin
,
T.
Berzina
, and
M. P.
Fontana
,
Cryst. Rep.
52
,
159
(
2007
).
8.
S. H.
Jo
,
T.
Chang
,
I.
Ebong
,
B. B.
Bhadviya
,
P.
Mazumder
, and
W.
Lu
,
Nano Lett.
10
,
1297
(
2010
).
9.
V.
Erokhin
,
T.
Berzina
, and
M. P.
Fontana
,
J. Appl. Phys.
97
,
064501
(
2005
).
10.
V.
Erokhin
,
T.
Berzina
,
P.
Camorani
,
A.
Smerieri
,
D.
Vavoulis
,
J.
Feng
, and
M. P.
Fontana
,
BioNanoScience
1
,
24
(
2011
).
11.
D. O.
Hebb
,
The Organization of Behavior
(
Wiley and Sons
,
New York
,
1961
).
12.
M.
Ziegler
,
R.
Soni
,
T.
Patelczyk
,
M.
Ignatov
,
T.
Bartsch
,
P.
Meuffels
, and
H.
Kohlstedt
,
Adv. Func. Mater
22
,
2744
(
2012
).
13.
F.
Alibart
,
S.
Pleutin
,
O.
Bichler
,
C.
Gamrat
,
T.
Serrano-Gotarredona
,
B.
Linares-Barranco
, and
D.
Vuillaume
,
Adv. Func. Mater.
22
,
609
(
2012
).
14.
M. N.
Awais
and
K. H.
Choi
,
Electr. Mater. Lett.
10
,
601
(
2014
).
15.
V.
Erokhin
,
T.
Berzina
,
K.
Gorshkov
,
P.
Camorani
,
A.
Pucci
,
L.
Ricci
,
G.
Ruggeri
,
R.
Sigala
, and
A.
Schuz
,
J. Mater. Chem.
22
,
22881
(
2012
).
16.
V.
Erokhin
,
T.
Berzina
,
P.
Camorani
, and
M. P.
Fontana
,
Soft Matter
2
,
870
(
2006
).
17.
S.
Gaba
,
P.
Sheridan
,
J.
Zhou
,
S.
Choi
, and
Z.
Lu
,
Nanoscale
5
,
5872
(
2013
).
18.
P.
Knag
,
W.
Lu
, and
Z.
Zhang
,
Ieee Trans. Nanotechnology
13
,
283
(
2014
).
19.
A.
Stotland
and
M.
Di Vetra
,
Phys. Rev. E
85
,
011116
(
2012
).
20.
G.A.
Patterson
,
P.I.
Fierens
,
A.A.
Garcia
, and
D.F.
Grosz
,
Phys. Rev. E
87
,
012128
(
2013
).
21.
J. H.
Cheung
,
W. B.
Stokton
, and
M. F.
Rubner
,
Macromolecules
30
,
2712
(
1997
).
22.
E. T.
Kang
,
K. G.
Neoh
, and
K. L.
Tan
,
Progr. Polymer Sci.
23
,
277
(
1998
).
23.
V. A.
Demin
,
V. V.
Erokhin
,
P. K.
Kashkarov
, and
M. V.
Kovalchuk
,
J. Appl. Phys.
116
,
064507
(
2014
).
24.
T.
Berzina
,
S.
Erokhina
,
P.
Camorani
,
O.
Konovalov
,
V.
Erokhin
, and
M. P.
Fontana
,
ACS Appl. Mater. Interfaces
1
,
2115
(
2009
).
25.
Y. V.
Pershin
,
S.
La Fontaine
, and
M.
Di Ventra
,
Phys. Rev. E
80
,
021926
(
2009
).
26.
K. D.
Cantley
,
A.
Subramanian
,
H. J.
Stiegler
,
R. A.
Chapman
, and
E. M.
Vogel
,
IEEE Trans. Nanotechnol.
10
,
1066
(
2011
).
27.
Z. Q.
Wang
,
H. Y.
Xu
,
X. H.
Li
,
H.
Yu
,
Y. C.
Lin
, and
X. J.
Zhu
,
Adv. Func. Mater.
22
,
2759
(
2012
).
28.
S. P.
Adhicari
,
C.
Yang
,
H.
Kim
, and
L. O.
Chua
,
IEEE Trans. Neural Networks Learning Syst.
23
,
1426
(
2012
).
29.
G.
Howard
,
E.
Gale
,
L.
Bull
,
B. D.
Costello
, and
A.
Adamatzky
,
IEEE Trans. Evolutionary Computation
16
,
711
(
2012
).
30.
S. Z.
Li
,
F.
Zeng
,
C.
Chen
,
H. Y.
Liu
,
G. S.
Tang
,
S.
Gao
,
C.
Song
,
Y. S.
Lin
,
F.
Pan
, and
D.
Guo
,
J. Mater. Chem. C
1
,
5292
(
2013
).
31.
L.
Chen
,
C. D.
Li
,
T. W.
Huang
,
Y. R.
Chen
, and
X.
Wang
,
Neural Computing Appl.
25
,
393
(
2014
).
32.
V.
Braitenberg
,
Vehicles: Experiments in Synthetic Psychology
(
MIT press
,
1984
).