You are on page 1of 14

1

A boundary condition-based approach to the modeling of memristor nano-structures.


Fernando Corinto, Senior Member, IEEE, and Alon Ascoli
AbstractA deep theoretical discussion proves that in Joglekars and Bioleks models the memductance-ux relation of a memristor driven by a sign-varying voltage source may only exhibit single-valuedness and multi-valuedness respectively. This manuscript derives a novel Boundary Condition-based Model for memristor nano-structures. Unlike previous models, the proposed one allows for closed-form solutions. More importantly, subject to the nonlinear behavior under exam, this model enables a suitable tuning of boundary conditions, which may result in the detection of both single-valued and multi-valued memductanceux relations under certain sign-varying inputs of interest. The large class of modeled dynamics include all behaviors reported in the legendary paper revealing the existence of memory-resistance at the nano-scale. Index Termsmemristor model, boundary conditions, nanodevices, nonlinear circuit

considering an oxide lm with opposite polarity, determines the occurrence of distinct dynamics. In the following, the dimensionless input-controlled length of conductive layer in the device with polarity coefcient set to 1 (-1) are denoted as w(t) D1 and 1 w(t) D1 respectively (D, the normalization factor, is the entire length of the thin oxide lm). Further, v(t) and i(t) respectively denote voltage across and current through the nano-structure. Finally, without loss of generality, we assume to control the device through a voltage source. The merit for the rst memristor model may be ascribed to Williams himself [3]. Linear [3]-[4] and nonlinear [10][12] dopant drift models fall into the following class (here dimensionless time variable = t t1 , where t0 is the time 0 normalization factor, is renamed as t for simplicity): dx(t) = W (x(t)) v(t) F (x(t), v(t), p) , dt i0 i(t) = W (x(t)) v(t), (1) (2)

I. I NTRODUCTION HE voltage- or current-controlled memristor is a fundamental two-terminal passive element with conductance depending on the time history of voltage across or current through it. It was theoretically introduced in 1971 [1] and then classied within the larger class of memristive systems [2]. However, it was only after the 2008 rst-ever recognition of memristor behavior at the nano-scale from Hewlett-Packard (HP) engineer Williams [3], [4] that it started to receive signicant attention for the large spectrum of opportunities it opens up in integrated circuit (IC) design, mostly for the realization of large-capacity non-volatile memories [5], [6] and neuromorphic systems [7], [8]. An optimal design of such ICs require a preliminary thorough study of their nonlinear dynamics [9]. With such a target in mind, it would be extremely helpful to develop a simple, accurate and general model capturing the behaviors of distinct memristor nanostructures. The HP nano-scale device (which we take as reference element for the class under modeling) is a thin lm of Titanium Dioxide (T iO2 ) sandwiched between two Platinum (P t) contacts [4]. The lm consists of a conductive layer of Oxygen-decient Titanium Dioxide (T iO2x , x = 5%) and of a insulating layer of stoichiometric Titanium Dioxide (T iO2 ). The lm conductivity may be suitably adjusted by applying an external input to the nano-device. Swapping the two layers, i.e.

where state variable x(t) represents the dimensionless length of conductive layer, i0 is a normalization factor1 , W (x(t)) denotes the memductance, expressed by W (x(t)) = Gon Gof f , Gon Gx(t) (3)

with G = Gon Gof f , where Gon and Gof f indicate the device memductance respectively in the fully-conductive and fully-insulating state (i.e. as x(t) sets to 1 and 0 respectively), while F (x(t), v(t), p) is a non-negative parameterized window function assuming values in [0, 1] and allowing one: to account for nonlinear effects on the ionic transport (p N+ modulates the degree of such nonlinearities); to impose system-dependent boundary conditions. The linear model, originally proposed in [3]-[4], assumes constant ionic drift rate throughout the lm length under application of an external input. The window function is thus dened as2 : F (x, v, p) = 1 x, v, p.

Copyright (c) 2012 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending an email to pubs-permissions@ieee.org. This work was partially supported by the Istituto Superiore Mario Boella and the regional government of Piedmont. F. Corinto and A. Ascoli are with the Department of Electronics, Politecnico di Torino, Torino, Italy, e-mail: (fernando.corinto@polito.it and alon.ascoli@polito.it).

This model, frequently adopted in literature [13]-[14] for its simplicity, qualitatively reproduces the dynamics of the memristor with a good degree of accuracy. However, since it does not specify boundary conditions, thus imposing no
1 Current i = q t1 is the magnitude of charge required to ow through 0 0 0 the device for the layer boundary to move across the entire lm length over time interval t0 . Charge q0 is given by q0 = D 2 Gon 1 , with standing for the average dopant mobility [3]. 2 In the following, unless it is strictly necessary, the explicit time dependence is dropped.

bounds on state x, it is valid only as long as the control source is chosen so as the layer interface does never reach any end. Literature was later enriched with a number of nonlinear dopant drift models. Those proposed by Williams himself [3], Joglekar [10] and Themis [11] introduce a decrease in rate of ionic transport as the layer boundary approaches any of the two ends. In particular, Joglekar proposed a window function let us call it FJ independent on v (see Fig. 2(a)): FJ (x, p) = 1 (2x 1)
2p

(a) : FJ , p = {1, 3, 10} (b) : FB , p = 2, = 1 1+


1 1

(c) : F , = 1

3 2 0

4 +
0.8 0.8

1+
0.8

4 3,1

6+,0

v,

(4)
FJ

0.6

0.6
FB

0.6
F

where p controls the window decrease rate (from the maximum unitary value at x = 0.5), occurring as x approaches either 0 or 1. For p = 1 (4) is identical (for less than a proportionality factor) to another window previously proposed by Williams himself in [3]. Furthermore, a slight modication of (4) was recently proposed in [11]. Due to the independence of (4) on the control source, Joglekars model may only allow for single-valued memductance-ux characteristics under any (and particularly under sign-varying) input. In Bioleks model [12] the dopant drift rate depends on a discontinuous window decreasing towards 0 as the layer boundary approaches any of the two ends and exhibiting vertical upward transitions each time source reverses polarity (see Fig. 2(b) as reference). Bioleks window function let us call it FB is expressed by { 1 x2p v > 0, (5a) FB (x, v, p) = 2p 1 (x 1) v 0. (5b) Due to its input-dependent expression, Bioleks model may only allow for multi-valued memductance-ux characteristics under sign-varying input. Summarizing, no model hitherto available in literature may reproduce both single-valued and multi-valued memductanceux characteristics under sign-varying input. The aim of this work is to ll such gap in literature. We propose a unique and accurate model we name it Boundary Condition-based Model (BCM) revolving around the linear dopant drift assumption, which simplies its analytical integration and endows it with a closed-form solution under any state initial condition/input combination. Particularly, this is shown for a periodic input with one change of polarity per period. Note that analytical solutions in closed form may not be extracted from Joglekars and Bioleks models. Most importantly, the BCM employs a novel window function see Fig. 2(c) as reference whose behavior at boundaries may be suitably tuned depending on the particular physical system under modeling, thus enabling the detection of the desired single-valued or multi-valued memductanceux characteristic under certain sign-varying inputs of interest. Particularly, under periodic zero-mean input with one change of polarity per period, necessary and sufcient conditions for single-valued memductance-ux relations are derived. This boundary behavior tuning feature allows the reproduction of a wide gamut of memristor behaviors recently observed in experiments on distinct nano-scale double-layer lms [17][19], including all the dynamical scenarios reported in the sensational paper on the rst measurement of memductance in nature [3].

0.4

0.4

0.4

p=1
0.2

p=3 p = 10

0.2

4+

20

0.2

5
1 0 0 0.5 x 1 0 0 0.5 x

2+
1

0.5 x

Figure 1. Window functions. Plot (a): Joglekars window for p = 1, 3, 10. Plot (b): Bioleks window for p = 2 and = 1. Plot (c): Proposed window for = 1. In plots (b) and (c) for = 1 arrows keep their directions, but curve number subscripts + and need to be replaced with and + respectively. Further, in plot (c) curve number subscripts , 1 and +, 0 need to be replaced with +, 1 (indicating v > vth,1 ) and , 0 (indicating v < vth,0 ) respectively.

As for the structure of the manuscript, Section II presents a deep theoretical analysis of Joglekars and Bioleks models. In particular, mathematical proofs for the single-valuedness and multi-valuedness arising in the memductance-ux relation under sign-varying input respectively in Joglekars and Bioleks models are given. Section III proposes the proposed model, proves the existence of a closed-form BCM solution under each state initial condition/input combination, presents a pseudo-code-based algorithm yielding the time interval over which each boundary condition holds under a periodic input with one change of polarity per period, and, provided such input also has zero-mean, determines the necessary and sufcient conditions for single-valued memductance-ux characteristics. Section IV is devoted to the BCM validation, including its use for modeling memristors employed in a chaotic oscillator. Finally, Section V outlines conclusions and future research developments. II. A NALYTICAL STUDY OF EXISTING MODELS In this section we deeply investigate the properties of the most important models hitherto available in literature so as to explain the reasons behind their inability to capture some of the dynamics reported in [3]. This analysis requires a brief preliminary description of the mechanisms underlying the dynamics of Joglekars and Bioleks nonlinear dopant drift models. A. Critical discussion of existing models Joglekars window function (4) is shown in Fig. 1(a). Under sign-varying input, its time evolution occurs along a single

curve. The window is traveled from left to right when x increases (i.e. for v > 0) and in the opposite direction when x decreases (i.e. for v < 0). As p gets bigger and bigger, the two x-values at which FJ decreases to 90% of its unitary value get farther and farther away from x = 0.5 and from such states the window decrease rate becomes larger and larger. This window is responsible for the existence of two equilibria, specically x = 0 and x = 1, in (1). Therefore, if state initial condition is set to either 0 or 1, then x keeps unaltered for all subsequent times, no matter what the input is. However, in a real case scenario, reversing the control source sign may cause x to change in the opposite direction. Due to the independence of (4) on the input, Joglekars model may only yield single-valued memductance-ux characteristics3 . This shall be demonstrated in Section II-B. Unlike (4), Bioleks window function (5a)-(5b) depends on the control source, i.e. it consists of a distinct branch per source polarity. Further, switching between these branches occurs through discontinuous vertical transitions. Let us consider a simple numerical example in order to clarify the time evolution of FB with variation in input v. Here we choose = 1 and p = 2. Fig. 1(b) displays (5a)-(5b) for this case. Setting x(0) = 0 and assuming an initially positive voltage, i.e. v(0) > 0, FB is expressed by (5a) and starts from 1. The positive voltage causes x to increase towards 1 according to (1). Correspondingly window function decreases with pdependent rate towards 0 along curve number 1+ . For claritys sake, in Fig. 1(b) the positive sign, negative sign and 0 value of the voltage associated to each numbered curve is respectively reported as subscript +, and 0 on the curve number. Assuming voltage sign is kept positive, at some instant x and FB become respectively equal to 1 and 0. Thereafter they shall maintain their respective values as long as voltage sign is kept unaltered. Note that for a negative or zero voltage it is (5b) which describes FB . Therefore, when voltage assumes the 0 value the window experiences a sudden transition from 0 to 1 (vertical solid line number 20 ). Obviously, voltage could set to 0 before that x takes the unitary value. If this occurs when x equals 0.9, FB jumps upwards along another curve, i.e. vertical dashed line 2 . 0 Assuming that afterwards voltage sign becomes negative, state variable shall decrease towards 0 with a rate set by (1). Correspondingly the window, expressed by (5b), decays towards 0 with pdependent rate along curve number 3 . As long as voltage sign is kept unaltered, at some time instant both x and FB shall equal 0. Provided that voltage sign is kept negative or voltage value gets equal to 0, no variation in the values of x and FB occurs thereafter. Only a new change in voltage sign produces a sudden 0-to-1 transition in the window (vertical solid line number 4+ ) since for a positive voltage FB is dened by (5a) once again. Certainly voltage sign could get positive before that x gets equal to 0. If this occurs when x
3 As it was observed in [20], a way to unequivocally dene a memristor is by specifying the charge-ux constitutive equation q = q() or the memductance-ux relationship W = W ().

equals 0.1, FB jumps upwards on another curve. i.e. vertical dashed line 4 . + Unlike Joglekars window, whose unitary maximum is at x = 0.5, Bioleks has unitary maxima at x = 0 and at x = 1. Keeping = 1, as p gets bigger and bigger, the x-value at which the window decreases to 90% of its unitary value with v > 0 (v < 0) gets farther and farther away from end x = 0 (x = 1) and from that state the window decrease rate becomes larger and larger. Contrary to Joglekars case, due to the dependence of (5a)(5b) on the input, Bioleks model may only allow for multivalued memductance-ux relations under sign-varying input. This shall be demonstrated in Section II-C. The next two sections report analytical integration of Joglekars and Bioleks models. This mathematical analysis where, without losing generality, is taken equal to 1 allows one to gain a deep insight into the memductance-ux relations arising from the models. In the numerical examples we consider a nano-structure with = +1, D = 10 nm, 2 = 1014 ms , Gon = 102 S and Gon = 60 Gof f . State V initial condition is set to x(0) = 0.1. Furthermore, the time normalization factor is set as the device characteristic time 1 t0 = D2 1 v0 [3], where v0 = 1 V denotes the amplitude of a sine-wave control voltage source, expressed by v(t) = v0 sin(2t), (6) with the normalized frequency, i.e. = f t0 R, set to 0.01, and f denoting the frequency in Hz. Finally, as it was done for time and length of conductive layer, in the simulation results we shall normalize all the physical quantities of interest: voltage v, ux , charge q and current i are respectively normalized with respect to v0 , 0 = v0 t0 , q0 and i0 = q0 t1 = Gon v0 . For the above specied param0 eter setting we have: t0 = 102 s (implying f = 1 Hz), 0 = 102 V s, q0 = 104 C and i0 = 102 A. B. Analytical integration of Joglekars equations Let us analytically solve for x(t) = x((t)) differential equation (1) using Joglekars window (4). Results shown in this section apply mutatis mutandis to the nonlinear drift models proposed in [3] and [11]. Note that in this case we do not need to specify the form of the input voltage. Mathematical calculations yield: x(t) 1 1 t dx(t ) = v(t ) dt . (7) i0 0 x(0) W (x(t )) FJ (x(t ), p) t Noting that (t) (0) = 0 v(t ) dt , letting z = 2x 1 (z (1, 1) being x (0, 1) ), using (3) and (4), (7) may be recast as
Gon + Gof f z(t) 1 G dz(t ) 4 Gon Gof f z(0) 1 (z(t ))2p 4 Gon Gof f z(t) (t) (0) z(t ) dz(t ) = . (8) 1 (z(t ))2p i0 z(0)

Before proceeding forward, we need to solve a couple of indenite integrals. Here we give the relative solutions:

1 dz 1 z 2p z dz 1 z 2p

= z =

1 1 , 1, 1 + , z 2p + k1 , (9) 2p 2p [ ] z2 1 1 F1 , 1, 1 + , z 2p + k2 , (10) 2 2 p p
2 F1

with k1 and k2 standing for integration constants. These solutions involve 2 F1 , denoting the Gauss Hypergeometric function [21], which, provided c N+ , converges to the following series
2 F1 [a, b, c, z]

(a)k (b)k z k k=0

In this simple case, using (15), it is straightforward to demonstrate that = (x) is a monotonic (increasing) function of x. As a result there exists a bijective mapping x = x(). This allows one to conclude that W = W () = W (x()), obtained by numerically nding the solution x = x() to (14) and inserting it into (3), is a single-valued function of . Similar conclusions may be drawn for p = 2. Numerical simulations suggest that this holds for each value of p (refer to Fig. 2). This is theoretically proved by the following proposition. Proposition 1: Equation (13) denes a bijective function x = x(). Proof: By recalling linear transformation z = 2 x 1, (13) may be rewritten as: [ ] 1 1 Gon + Gof f z F1 , 1, 1 + , z 2p 2 Gon Gof f 4 2p 2p [ ] 2 G z 1 2p 1 , 1, 1 + , z 2 F1 Gon Gof f 8 p p (t) (0) + , (16) = i0 where is a constant accounting for all terms related to initial condition z(0) = 2x(0) 1 in (13). Mapping x = x() is bijective if and only if = (z), dened by (16), is monotonic, i.e. if (z) = d (z) does not change sign for dz all z (1, 1). Some simple algebraic manipulations permit to express the contiguous Gauss Hypergeometric functions appearing in (16) as power series: [ ] + 1 1 1 F1 , 1, 1 + , z 2p = z 2pk , (17) 2 2p 2p 1 + 2pk k=0 [ ] + 1 1 1 , 1, 1 + , z 2p = z 2pk . (18) 2 F1 p p 1 + pk
k=0

(c)k k!

(11)

for |z| < 1. Further, (x)k denotes the Pochhammer symbol, dened as (x)k = (x + k) = x(x + 1) (x + k 1), (x) (12)

with , the Gamma function, related to the factorial through (x) = (x 1)! and k N0 . Using (9)-(10), (8) may be analytically solved for each value of p and for each input v(t). The relationship between x(t) and (t) is found to be4 :
( ) Gon + Gof f 1 1 x(t) Gon Gof f 2 2 [ ( 1 1 p , 1, 1 + , 4 x(t) 2 F1 2p 2p [ ( )2 1 1 1 x(t) , 1, 1 + 2 F1 2 2 p

G Gon Gof f ( )2p ] 1 p 1 , 4 x(t) p 2 ( ) (t) (0) Gon + Gof f 1 1 = + x(0) i0 Gon Gof f 2 2 [ ( )2p ] 1 1 p 1 G , 1, 1 + , 4 x(0) 2 F1 2p 2p 2 Gon Gof f [ ( )2 ( )2p ] 1 1 1 1 p 1 x(0) , 1, 1 + , 4 x(0) (13) . 2 F1 2 2 p p 2

1 2

)2p ]

These expressions allow one to rewrite (16) as it follows:


(t) (0) + i0 =
+ Gon + Gof f 1 z 1+2pk 4 Gon Gof f 1 + 2pk k=0 + 1 G z 2(1+pk) . (19) 8 Gon Gof f 1 + pk k=0

It is not possible to determine a closed-form expression for x = x() from (13), not even for p = 1, when (13) gets a simpler expression: ( )Gon ( )G x(t) 1 x(0) of f x(0) 1 x(t) ) ( (t) (0) . (14) = exp 4 Gon Gof f i0 Equation (14) denes the following function (t): i0 (t) = (x(t)) = (0) + 4 Gon Gof f ) ( )) ( ( G G (x(0)) on (x(t)) on ln (15) . ln G G (1 x(t)) of f (1 x(0)) of f
4 A similar calculation for the solution to the class of Abel differential equations consisting of all possible state equations derived from (1) with Joglekars window function (4) and p swept in N, is also shown in [15].

Equation (19) enables the study of monotonicity of = (z) through evaluation of the sign of its rst-order derivative, which, unlike proportionality factor i0 (positive by denition), is expressed by:
1 (z) i0
+ + 2pk+1 Gon + Gof f 2pk G z z 4 Gon Gof f 4 Gon Gof f k=0 k=0 ( ) + Gon + Gof f G z z 2pk 4 Gon Gof f 4 Gon Gof f k=0 ( ) 1 Gon + Gof f z G 1 , (20) 4 Gon Gof f 1 z 2p

= =

FJ , p 1
W 0.0010 0.0008 0.0006 0.0004 0.0002 0.0000 0.0002 0.0004
Figure 2. Single-valued memductance-ux characteristics for the memristor modeled by (1)-(2), with boundary conditions dened through Joglekars window (4) where p = {1, 2, 10} and input given by (6). For such input, in (13) we have (t) (0) = v0 1 1 sin2 (t).

i i0 1

FJ , p 1

FJ , p 2
FJ , p 2

0.5

FJ , p 10
FJ , p 10

1.0
0 1

0.5

0.5

1.0

v v0 1

500

1000 1500 2000 2500 3000 3500

0.5

Figure 3. Plots of i = W () v versus v corresponding to the W ()- relations in Fig. 2.

where in the last expression we used the convergence property of the Geometric series for |z| < 1. Equation (20) implies that = (z) is monotonic if function (z) = Gon + Gof f z G (21)

does not change sign z (1, 1). It is straightforward to realize that (z) is positive z (1, 1). Thereby, the monotonicity of (z) is demonstrated. This proves that x = x(), given by (13), is single-valued. Since x may not be expressed as function of by means of (13), a closed-form expression for W = W () = W (x()) may not be given either. It follows that the memductance-ux characteristic may only be determined through numerical simulations, as it was suggested for p = 1. Anyway, Proposition 1 allows one to conclude that W = W () = W (x()) is a single-valued function of . Using the parameter setting from Section II-A, numerically solving (13) for x = x() and inserting this solution into (3), Fig. 2 shows the single-valued memductance-ux characteristics obtained from Joglekars model for p respectively set to 1 (blue curve), 2 (black curve) and 10 (red curve). Despite the shapes of such characteristics are independent of the input source, the abscissa and ordinate values for each curve in Fig. 2 refer to the previously dened sine-wave voltage driving. The current-voltage characteristics corresponding to the memductance-ux relations in Fig. 2 are shown in Fig. 3, where the blue, black and red curve respectively refer to p equal to 1, 2 and 10. It is worth noting that the red curve in Fig. 3 qualitatively reproduces Fig. 3(c) from [3], referring to a dynamical scenario characterized by strong nonlinearities in ionic transport due to the large electric eld developing within the conductive layer of the HP thin oxide lm under

sine-wave control voltage source. However, none of the iv characteristics of Fig. 3 resembles Fig. 3(a) from [3], referring to a dynamical scenario where the memristor exhibits an apparent or dynamical negative differential conductance. We ascribe this modeling inability to the single-valuedness of x = x(), previously demonstrated in Proposition 1 and shown in Fig. 2. This may be summarized in the following result: Result 1: Joglekars model may only yield single-valued memductance-ux characteristics. By contrast, under sign-varying input, Bioleks model may only allow for multi-valued memductance-ux characteristics. Theoretical demonstration for this is given in the next section, where it is also shown that, under the same parameter setting and driving condition of Figs. 2-3, Bioleks model detects Fig. 3(a) from [3], but fails to capture Fig. 3(c) from [3].

C. Analytical integration of Bioleks equations State equation (1) with Bioleks window (5a)-(5b) may only be integrated provided some information about the control voltage source is known, since Bioleks window depends upon its sign. We are interested in typical input signals with periodicity, some polarity change over the period and zero mean (examples include square-waves, sawtooth-waves and sine-waves such as (6)). For simplicity we assume that the source changes its sign just once over the period T . Without loss of generality, we assume that v(t) is positive for t (0, t), , T ) and has zero mean, i.e. negative for t (t { v(t) > 0 v(t) 0 t (0, t), t [t, T ], (22a) (22b)

Integrating (1) for t (0, t) since v(t) > 0, (5a) is the branch of FB under study yields: x(t) 1 1 1 t dx(t ) = v(t ) dt . (23) 2p i0 0 x(0) W (x(t )) 1 (x(t )) Using (3) and (9)-(10), (23) may be analytically solved for each value of p and for each input v(t) satisfying (22a). The relationship between x(t) and (t), t (0, t), is given by:
] 1 1 , 1, 1 + , (x(t))2p Gof f 2p 2p [ ] G 1 1 1 2 (x(t)) 2 F1 , 1, 1 + , (x(t))2p Gon Gof f 2 p p ] [ (t) (0) 1 1 1 = , 1, 1 + , (x(0))2p + x(0) 2 F1 i0 Gof f 2p 2p [ ] G 1 1 1 2 2p (x(0)) 2 F1 , 1, 1 + , (x(0)) , (24) Gon Gof f 2 p p 1 x(t) 2 F1 [

t where (t) (0) = 0 v(t )dt . By contrast, integrating (1) for t [t, T ] since v(t) 0, (5b) is the branch of FB under study lead to:

x(t) x(t)

for all t (0, t) and through 1 ( ( ) )G G x(t) 2 x(t) on 2 2 x(t) 2 x(t) x(t) 2 x(t) ) ( (t) (t) (29) = exp Gon Gof f i0 for all t [t, T ]. It is possible to demonstrate that equations (28) and (29) represent distinct bijective functions x = x(), respectively valid for v(t) > 0 and v(t) 0. It follows that mapping x = x() dened by (28)-(29) loses bijectivity as v(t) changes polarity and vertical transitions see Fig. 1(b) in Section II-A lead Bioleks window to evolve from one branch to the other. Numerically solving (28) under t (0, t) , T ] for x(t) = x((t)) and then inserting and (29) under t [t these solutions into (3) yields a multi-valued memductance dependence on ux for t [0, T ]. The multi-valuedness of function x = x() expressed by (24)-(27) and thus of the corresponding memductance-ux mapping W () = W (x()) may be proved for each value of p, as it is reported in the following proposition. Proposition 2: Under sign-varying control voltage source, equations (24)-(27) dene a multi-valued x = x() function. Proof: Equations (24) and (27) may be rewritten in a form similar to (16). In particular, (24) with z = x (0, 1) is [ ] 1 1 1 2p z 2 F1 , 1, 1 + , z + Gof f 2p 2p [ ] G 1 1 2p 2 z 2 F1 , 1, 1 + , z = 2 Gon Gof f p p (t) (0) = + + , (30) i0 while (27) with z = x 1 (1, 0) turns into [ ] 1 1 1 2p z 2 F1 , 1, 1 + , z + Gon 2p 2p [ ] G 1 1 z 2 2 F1 , 1, 1 + , z 2p = 2 Gon Gof f p p ) (t) (t = + , (31) i0 where + and are constants accounting for all terms related to initial conditions z(0) = x(0) in (24) and z(t) = ) 1 in (27) respectively. Using (17)-(18), from (30) it is x(t readily derived that Gon z G 1 1 d (z) = > 0 z (0, 1). i0 d z Gon Gof f 1 z 2p In the same way, from (31) it follows that 1 d (z) Gof f z G 1 = > 0 z (1, 0). (33) i0 d z Gon Gof f 1 z 2p Thereby, under positive (non-positive) input, the monotonicity of = (z) follows from (32) ((33)). This proves the singlevaluedness of x = x() numerically derivable from (24) ((27)) for t (0, t) (t [t, T ]). It follows that a sign change in v(t) may yield a couple of distinct values of x for a given value of . Thus, under sign-varying control voltage source, (24)-(27) dene a non-bijective x = x() mapping. (32)

1 1 1 dx(t ) = W (x(t )) 1 (x(t ) 1)2p i0

v(t ) dt . (25)

t Noting that t v(t ) dt = (t) (t), letting z = x 1 and using (3), (25) may be recast as
z(t) 1 1 G dz(t ) Gon z(t) 1 (z(t ))2p Gon Gof f z(t) z(t ) (t) (t) dz(t ) = . 2p i0 ) 1 (z(t )) z(t

(26)

Using (9)-(10), (26) may be analytically solved for each value of p and for each input v(t) satisfying (22b). The relationship between x(t) and (t), t [t, T ], is given by:
[ ] 1 1 1 (x(t) 1) 2 F1 , 1, 1 + , (x(t) 1)2p Gon 2p 2p [ ] G 1 1 1 2 2p (x(t) 1) 2 F1 , 1, 1 + , (x(t) 1) Gon Gof f 2 p p (t) (t) = + i0 [ ] 1 1 1 + (x(t) 1) 2 F1 , 1, 1 + , (x(t) 1)2p + Gon 2p 2p [ ] 1 1 G 1 2 (x(t) 1) 2 F1 , 1, 1 + , (x(t) 1)2p . Gon Gof f 2 p p (27)

Similarly to Joglekars case see (15) (24) and (27) may be recast in simpler forms for p = 1, when x(t) and (t) are related through )Gon ( )Gof f 1 ( 2 1 + x(t) 1 (x(0))2 2 1 + x(0) 1 (x(t)) ) ( (t) (0) (28) = exp Gon Gof f i0

FB , p 1

i i0 1 0.6 0.4 0.2 0.5 0.2 0.4 0.6 0.5 1.0 v v0 1

W 0.010

FB , p 1

FB , p 2

FB , p 2

FB , p 10

0.005 0.000 0.005

FB , p 10

1.0
0 1

500 1000 1500 2000 2500 3000 3500

Figure 4. Plots of W () = W (x()) for the memristor modeled by (1)-(2), with boundary conditions dened through Bioleks window (5a)-(5b), where p = {1, 2, 10}, and input given by (6). For t (0, t) in (24) we have (t) (0) = v0 1 1 sin2 (t), while for t [t, T ] in (27) we have (t) (t) = v0 1 1 cos2 (t) (dimensionless time period is T = 1 and t = 0.5 T ).

Figure 5. Current-voltage behaviors corresponding to the memductance-ux relations of Fig. 4.

Under the same parameter setting and driving condition described in Section II-A, numerically solving for x = x() (24) as v(t) > 0 and (27) as v(t) 0, and then inserting the corresponding solutions into (3), Fig. 4 depicts the multivalued W () characteristics obtained from Bioleks model with p equal to 1 (blue curve), 2 (black curve) and 10 (red curve). Note that, although the shapes of such characteristics do only depend on the sign of the input, the abscissa and ordinate values for each curve in Fig. 4 refer to the input specied in (6). The current-voltage characteristics corresponding to the memductance-ux plots of Fig. 4 are depicted in Fig. 5, where the blue, black and red curves respectively refer to p equal to 1, 2 and 10. The red curve resembles Fig. 3(a) from [3]. However, none of the i-v characteristics of Fig. 5 captures Fig. 3(c) from [3]. We ascribe this modeling inability to the multi-valuedness of x = x(), previously demonstrated in Proposition 2 and shown in Fig. 4. This may be summarized in the following result: Result 2: Under sign-varying control voltage source, Bioleks model may only yield multi-valued memductance-ux characteristics. Given Result 1 and Result 2, the aim of the remaining part of the manuscript is to derive the simplest possible analytical model able to characterize both single-valued and multi-valued memductance-ux characteristics under certain sign-varying inputs of interest. The simplicity of the proposed model, ensured by the linear dopant drift assumption, allows it to possess closed-form solutions, i.e. it is possible to explicitly determine x(t) = x((t)) for all t under each state initial condition/input combination. No other model exhibits such

attractive feature. The proposed linear dopant drift-based model is then endowed with boundary condition tuning capability, which, depending on the system under consideration, permits the detection of the desired single-valued or multi-valued memductance-ux characteristic under certain sign-varying inputs of interest. The derivation of the proposed model was inspired by the work of Williams et al., who, in [3], stated that the switching characteristic observed for a particular memristive system helps to classify the nature of the boundary conditions on the state variable of the device. Next section shall introduce the proposed model. III. B OUNDARY CONDITION - BASED MODEL FOR MEMRISTOR NANO - STRUCTURES In this section we present a novel Boundary Condition-based Model (BCM) for memristor nano-structures. We assume that ionic drift rate under an external input is constant throughout the lm extension. This provides a simple model which allows for closed-form solutions under any state initial condition/input combination. Further, dynamics at boundaries may be suitably tuned so as to capture a wide range of nonlinear behaviors from various distinct nano-structures under certain sign-varying inputs of interest. The proposed model is based on a window function having unitary value for all x(t) (0, 1) and exhibiting the following vertical transitions (see Fig. 1(c)): { if t | x(t) = 1 for v(t) vth,1 , (34a) 10 if t | x(t) = 0 for v(t) vth,0 , (34b) and 01 { if t |x(t) = 1 if t |x(t) = 0 for v(t) < vth,1 , (35a) for v(t) > vth,0 . (35b)

The occurrence of such vertical transitions depend on the values of non-negative parameters vth,1 and vth,0 . Conditions (34a) and (34b) respectively establish when x(t) sticks to boundaries x = 1 and x = 0. By contrast, conditions (35a) and (35b) dene when x(t) leaves these boundaries. A mathematical description of these dynamics is obtained by dening tunable boundary conditions Cn (n = 1, 2, 3) as C1 C2 C3 = = = { x(t) (0, 1) or (x(t) = 0 and v(t) > vth,0 ) or (x(t) = 1 and v(t) < vth,1 )}, (36) { x(t) = 0 and v(t) vth,0 }, { x(t) = 1 and v(t) vth,1 }, (37) (38)

where, referring to the case = +1 for simplicity, vth,0 (vth,1 ) denotes the threshold voltage the magnitude of the input needs to pass over after it gets positive (negative) while (37) ((38)) holds, before (36) may be met. Thereby, the mathematical expression for the proposed window function F results to be: { 1 if (36) holds, F (x, v) = (39) 0 if (37) or (38) holds. In the matched case, when threshold voltages are equal, condition vth,0 = vth,1 = vth is satised. A brief qualitative description of the proposed BCM is worthwhile prior to its analytical integration. A. Qualitative description As in Bioleks case, the proposed window depends on the control source. Under sign-varying input, for x (0, 1) it may evolve on a single curve only (similarly to Joglekars case), while for x at boundary x = 1 (x = 0), it may experience the 1 0 and 0 1 vertical transitions specied by (34a) and (35a) ((34b) and (35b)) respectively. Taking Fig. 1(c) as reference, let us clarify these dynamics through a numerical example, where is set to +1. Setting x(0) = 0 and v(0) > vth,0 , condition (36) is fullled, thus (39) implies an initial unitary value for our window. Assuming that afterwards voltage sign is positive, x increases towards 1 according to (1), while the window keeps its unitary value (see line 1+ ). For claritys sake, in Fig. 1(c) scenarios with v(t) > 0, v(t) < 0, v(t) > vth,0 and v(t) < vth,1 associated to each numbered line are respectively reported as subscripts +, , +, 0 and , 1 on the line number. Provided voltage sign is kept positive, at some point x gets equal to 1, implying fulllment to condition (38) and causing window (39) to experience the sudden 1 0 transition dened in (34a) (see line 2+ ). After that, state variable and window function keep their respective values unless voltage sign is reversed and voltage magnitude gets larger than vth,1 . In fact, in case this happens, condition (36) holds once again and window (39) experiences the sudden 0 1 transition dened in (35a) (see line 3,1 ). Later, provided voltage sign is kept negative, solution x to (1) decreases towards 0 while the window keeps its unitary value (see line 4 ). Provided voltage sign is kept negative, at some point x gets equal to 0, condition (37) is satised, and the sudden 1 0 transition, dened in (34b) (see line 5 ), is

triggered in window (39). Then both x and F keep equal to 0 unless voltage sign is reversed and voltage magnitude gets larger than vth,0 once again. If this occurs, (36) is met yet again and the window experiences the sudden 0 1 transition dened in (35b) (see line 6+,0 ). The main difference with Bioleks window (see Fig. 1(b)) lies in the fact that ours may only experience vertical transitions when x equals 0 or 15 . In fact, in case voltage sign is reversed while x (0, 1), this simply causes state variable to vary in the opposite direction according to (1) (with (39) keeping its unitary value due to the fulllment to condition (36)) thereafter. In this case, for = 1, the window time evolution occurs only along one path, dened by lines 1+ and 4 for positive and negative v(t) respectively (see Fig. 1(c)). This qualitative description is instrumental to a clear understanding of the analytical integration of the proposed model, presented in the next section. B. Analytical integration and closed-form solution A deep theoretical analysis of the proposed model is needed in order to highlight its most prominent capabilities, i.e. the ease of analytical integration, and, more importantly, the existence of closed-form solutions and the attractive property of tunable boundary conditions which makes it possible to observe both single-valued and multi-valued memductance-ux characteristics under certain sign-varying inputs of interest. Without losing generality, we take = 1. Assuming that at some instant ti condition C1 is met, let us analytically integrate (1) using the proposed window (39) under boundary condition (36), i.e. x(t) 1 t 1 dx(t ) = v(t ) dt , (40) i0 ti x(ti ) W (x(t )) where x(ti ) is the state initial condition. Some straightforward algebraic steps yield the following closed-form solution x(t) to (40) for all t such that condition (36) is satised: Gon G ( Gon G )2 + (x(ti ))2 2Gon x(ti )+ G (41)

x(t) =

2 Gon Gof f

t where (t) (ti ) = ti v(t )dt . Analytical integration of (1) using the proposed window (39) under boundary condition (37) and (38) is obvious, respectively yielding x(t) = 0 t | C2 holds, x(t) = 1 t | C 3 holds. (42) (43)

(t) (ti ) , G i0

5 In one of its forms, particularly in the matched case with v th,0 = vth,1 = vth = 0 V , (39) is an approximation to Bioleks window for very large p (compare Figs. 4-5 (red curves) to Figs. 8-9 (blue curves) and also refer to Section IV-C). The vertical transitions of Fig. 1(c) do not realistically reect physical phenomena. However, under parameter setting and driving condition yielding BCM solutions allowing for vertical transitions in F , various other memristor behaviors, besides the HP device dynamics of Figs. 8-9, may be qualitatively reproduced (see Section IV-B).

As in the case of Bioleks model (see Section II-C), before being able to characterize the time dynamics of x(t) in more detail, we need to specify some information on the control source. In line with the hypothesis in the discussion of Bioleks model, we assume that the input is given by (22a)(22b). The time interval over which each of boundary conditions (36)-(38) holds may be obtained by means of Algorithm 1, written in pseudo-code and based upon the discussion in Section III-A. The algorithm is started up with ti set to 0 thus x(ti ) = x(0) is the initial condition on state x(t) and then continuously run over a certain time interval [0, tend ]. Algorithm 1 Pseudo-code-based algorithm for the determination of the time interval over which each boundary condition holds. 1: procedure BCM 2: Given (22a)(22b) 3: Given vth,0 , vth,1 4: Given tend 5: x(0) [0, 1] Set x(0) as 0, 1, or in (0, 1) 6: while x(t) (0, 1) do C1 holds 7: x(t) (41) 8: end while 9: if x(t) = 1 then 10: t1 t 11: while v(t) vth,1 do C3 holds 12: x(t) 1 13: end while 14: t t 15: go to 6 16: else 17: t0 t 18: while v(t) vth,0 do C2 holds 19: x(t) 0 20: end while 21: t t 22: go to 6 23: end if 24: ts tend T 25: Return t1 , t , t , t0 [ts , tend ] 26: end procedure Without loss of generality we assume that tend is an integer multiple of T . Let us focus on the last period, i.e. ts dened on line 24 of the Algorithm 1 is assumed as reference instant and is reset to 0 for simplicity (implying tend = T ). Algorithm 1 permits a clear identication of time interval [t1 , t ] with t1 [0, t) and t [t, T ) when x((t)) halts at 1 with v(t) vth,1 and of time interval [0, t ] [t0 , T ] with t [0, t) and t0 [t, T ) when x((t)) halts at 0 with v(t) vth,0 . As a result, using (41)-(43), a closed-form solution x(t) = x((t)) to the proposed BCM may be obtained for each t. C. Conditions for single-valuedness of x = x() under certain sign-varying inputs of interest The following proposition presents the conditions ensuring the single-valuedness of the state-ux characteristic (and,

consequently, also of the memductance-ux characteristic) observed in the proposed BCM under certain sign-varying inputs of interest. Proposition 3: Given v(t) as in (22a)-(22b) with the following additional condition: T v(t )dt = 0. (44)
0

The BCM state-ux relationship x(t) = x((t)) is singlevalued if the following conditions are both fullled: t v(t )dt = 0 (45)
t1

v(t )dt = 0

(46)

t0

Proof: Let us take x(0) (0, 1) and hypothesize that, within a period, x(t) gets only stuck to 1 for t [t1 , t ]. Let us derive the necessary and sufcient conditions under which x(t) = x((t)) is single-valued. The proof is readily derived by means of the following steps based on Algorithm 1: 1) The initial point of the trajectory on the (, x()) plane is O = ((0), x(0)). 2) State x(t) increases according to (41) for all t [0, t1 ], where t1 [0, t) is the rst instant at which x(t) = 1. Due to (22a), during the same time interval (t) increases with time assuming values in [(0), (t1 )], t where (t1 ) = (0) + 0 1 v(t )dt . 3) At t = t1 , trajectory on the (, x()) plane reaches point P = ((t1 ), 1). 4) Due to (22a), for all t (t1 , t ] ux increases with time, assuming values in ((t1 ), (t)]. 5) With t = t [t, T ) denoting the rst time at which v(t) < vth,1 and due to (22b), for all t (t, t ] ux decreases with time, assuming values in [(t ), (t)), t where (t ) = (0) + 0 v(t )dt . Note that for all t [t1 , t ] state x(t) halts at 1. 6) At t = t , trajectory on the (, x()) plane reaches point Q = ((t ), 1). 7) State x(t) decreases according to (41) for all t (t , T ]. Due to (22b), during the same time interval (t) decreases with time assuming values in [(T ), (t )). It turns out that on the (, x()) plane trajectory evolves along the same curve (but with opposite direction) as at step 1 if and only if P Q. Further, since (44) implies (T ) = (0) and consequently x(0) = x(T ), the ending point of such trajectory is point O = ((0), x(0)). As a result, x = x() is single-valued. Condition P Q from the last step implies the fulllment to condition (45). The scenario where x(0) (0, 1) and, within a period, x(t) gets only stuck to 0 for t [0, t ] [t0 , T ] allows one to determine, mutatis mutandis, condition (46). Scenarios with different choice for x(0) and/or x(t) getting stuck to both boundaries within a period yield the same conditions (45)-(46) for single-valuedness in the state-ux characteristic under input (22a)-(22b) with (44) (their analysis

10

Table I C APABILITY OF EMULATION OF VARIOUS MEMRISTOR BEHAVIORS FROM [3] FOR VARIOUS MODELS PROPOSED IN LITERATURE . Memristor i v response from [3] Fig. Fig. Fig. Fig. Fig. 2(b) in [3] 2(c) in [3] 3(a) in [3] 3(b) in [3] 3(c) in [3] Linear Joglekars Bioleks BCM Fig. 6 Fig. 7 Fig. 9 (blue) Fig. 9 (black) Fig. 9 (red)

0.15 0.1 0.05


i i1 0
x

0.05 0.1 0.15 0.2 1 0.5

W ()

    

    

    

0.95

(a)
0.9 990 992 994

x 10 2

996

998 1000

1.5

follows the same guidelines of the proof of Proposition 3 and is thus omitted here). Condition (45) ((46)) establishes that x(t) = x((t)) (and, as a consequence, also W (t) = W ((t))) is single-valued if and only if the net ux (t) over time interval [t1 , t ] (t [0, t ] [t0 , T ]) is equal to 0. Remark 1: Conditions (45) and (46) are similar to conditions for zero net ux injection over a certain time interval [16]. The fulllment of such conditions is necessary in memristor-based memories to avoid corruption of memristor state after read cycles. To the best of our knowledge no quantitative result has ever been provided in literature to relate single-valuedness in the memductance-ux characteristic to the zero net ux injection property. In the next section we show how our model may describe various dynamics observed in different memristor nano-scale lms, including all behaviors from [3]. IV. VALIDATION OF THE PROPOSED MODEL This section is devoted to the BCM validation and consists of three parts, respectively focusing on dynamics reported in [3] (Section IV-A), on behaviors from a number of distinct physical systems (Section IV-B) and on chaos in a memristorbased oscillator (Section IV-C). A. Modeling the i-v response of HP nano-device In Table I we summarize the ability of detection of dynamics reported in [3] for various models. The second column was drawn according to [3], the second and third ones report comments from [12], while the last one shows simulation results discussed in this section. Unless stated otherwise, in the following numerical simulations we choose , D, and Gon as in Section II-A. Further, the adopted memristor model is described by (1)-(2) with window (39) tuned through boundary conditions (36)-(38). In this section we show how the proposed model may qualitatively reproduce all behaviors from [3]. Memristor behavior of Fig. 2(b) in [3] In Fig. 6 we numerically simulate our model through an ordinary differential equation solver. The on/off conductance ratio is Gon G1f = 160. The system is excited by control of source (6) with v0 = 1 V . State initial condition is taken as x(0) = 0.94. In boundary conditions (36)-(38) we specify vth,0 = vth,1 = vth = 0 V . The current-voltage plots (the

1 0 10

(b) 1 0
20 30

1 v v0

0.5

Figure 6. Memristor i-v response to sine-wave excitation source (6) with v0 = 1 V and equal to 1 (blue) and 10 (black). As it is expected of a memristor, v(t) and i(t) exhibit zeros at same times. Further, in each cycle maximum (minimum) in v(t) occurs before (after) maximum (minimum) in i(t). Insets (relative to the = 1 simulation): time plot of x (a) and memductance-ux characteristic (b).

blue and black curves, respectively associated to = 1, i.e. f = 100 Hz, and = 10, i.e. f = 1000 Hz, show the collapse of the bow-tie-shaped hysteretic loop to a straight line with a tenfold increase in frequency, a distinctive feature of a memristor) of Fig. 6 resemble Fig. 2(b) from [3]. The blue arrows show the direction of the loop relative to the excitation source at frequency f = 100 Hz. The insets (a) and (b) of Fig. 6, relative to the simulation with f = 100 Hz, respectively visualize the time plot of the oscillating state variable x(t) and the single-valued memductance-ux characteristic. Here x keeps away from the two boundaries for all t, thus conditions for single-valuedness of x = x() are obviously fullled. Memristor behavior of Fig. 2(c) in [3] Let us set state initial condition x(0) to 0.87. The system has Gon G1f = 80 and is controlled through a different periodic of k= sign-varying input, i.e. v(t) = k= v (tkT ) where T = 3, v (t) is dened as v0 sin2 (2t)[h(t t) h(t t )] for t [0, T ] and 0 otherwise, v0 = 1 V , = 1 i.e. f = 100 Hz, and h(t) refers to the Heaviside function, equal to 1, 0.5 and 0 for t > 0, t = 0 and t < 0 respectively (t = 1.5). As in the previous scenario, we use an ode solver to numerically integrate the proposed model and take zero matched threshold voltages in boundary conditions (36)-(38). The resulting current-voltage characteristic in Fig. 7 looks like Fig. 2(c) in [3]. System state x is at all times within open interval (0, 1), therefore a soft-switching scenario occurs6 and

6 By soft-switching scenario we mean a case where condition (36) is satised at all times and single-valued memductance-ux relations are observed. By contrast, by hard-switching scenario we denote a case where at least one of conditions (37) and (38) is met for some time. In this scenario multi-valuedness arises in the memductance-ux characteristics unless the occurrence of suitable conditions for single-valuedness, such as (45)-(46) derived in Section III-C for certain sign-varying inputs of interest, is ensured.

11

0.3
W ()

0.01

(a)

3 2

Fig. 3 a in 3 W 0.010

Fig. 3 b in 3 P' Q' P' P'Q'

Fig. 3 c in 3

0.2 0.1
i i1 0

0.005

0 0

20

40

60

80

Q'

0 0.1 0.2
4 6 5

1 0
1 0.5
1 v v0

0.4 0.2

0.005
1 i i0

0 0.5 1 0

4 1 (b)
1

2 3

0.3 0.4 1 0.5

0.2

0.000

0.4 3

1000 2000 3000 4000 5000 6000

0 1

0
1 v v0

0.5

0.005
Figure 8. Memductance-ux characteristics for the memristor modeled by (1)-(2) (x(0) = 0.1, Gon = 102 S), using proposed window (39) and voltage source (6) with = 0.01. Blue curve: v0 = 1 V , Gon G1f = 60, of vth = 0 V (note the similarity to red curve in Fig. 4). Black curve: v0 = 2 V , 1 1 Gon Gof f = 25, vth = 0 V . Red curve: v0 = 1 V , Gon Gof f = 60, vth = 0.5v0 V (note the similarity to red curve in Fig. 2). For each case, points P = ((t,1 ), Gon ) and Q = ((t ), Gon ) are highlighted to show whether or not one of conditions for single-valuedness, i.e. (45), is fullled. Here P Q for the red curve only.

Figure 7. Multiple loops the memristor i-v characteristic due to periodic in sign-varying input v(t) = k= v (t kT ), where T = 3, v is dened k= as v0 sin2 (2t)[h(t t) h(t t )] for t [0, T ] and 0 otherwise, v0 = 1 V , = 1, h(t) refers to the Heaviside function and t = 1.5. Inset (a): W ()- characteristic. Inset (b): Time waveforms of v(t) (blue) and i(t) (black).

the memductance-ux characteristic is single-valued, as it is shown in inset (a). Blue and black curves in inset (b) depict voltage and current waveforms respectively. Running the simulation up till t = T , the consecutive voltage and current 0.5T -wide lobes in inset (b) and the corresponding loops on the i-v plot are numbered from 1 to 6 (note that T = 1 = 1). In the following three scenarios we use rst pseudo-codebased Algorithm 1 from Section III-B to derive the time intervals where each of boundary conditions Cn (n = {1, 2, 3}) is fullled and then (41), (42) and (43) to determine the closedform solution to the BCM. v0 For the given input choice we have (t) = 2 cos(2t) v0 and (ti ) = 2 cos(2ti ). Using (41), the closed-form expressions for instants t, ( = {0, 1}), t and t are found to be: ( 1 i0 t, = arccos cos(2ti ) 2 v0 Gof f ( )) G 2 (x (ti ) ) + 2( x(ti )) , (47) Gon ( ) 1 vth,1 t = arcsin , (48) 2 v0 ( ) 1 vth,0 t = arcsin . (49) 2 v0 In (47) ti denotes the initial time of the interval [ti , t, ] over which C1 is fullled. Memristor behavior of Fig. 3(a) in [3] Let us reconsider sine wave voltage source (6). Choosing v0 , , the on/off conductance ratio Gon G1f and the state of

initial condition x(0) as in the simulations from Section II, the BCM detects the hard-switching scenario reported in Fig. 3a of [3] for vth,0 = vth,1 = vth = 0 V in (36)-(38). In this case x temporarily gets stuck to both boundaries over period T . In 8 (blue curve) we observe a multi-valued memductance-ux characteristic, since the earlier specied conditions for single-valuedness, i.e. (45)-(46) (see Section (III-C)), are not satised. This is demonstrated for (45) by the fact that P = ((t,1 ), Gon ) and Q = ((t ), Gon ) (blue points) do not overlap on the W () characteristic. As a result, as it was observed in [12], these dynamics may not be captured by Joglekars model (see Table I). The current-voltage characteristic is shown in Fig. 9 (blue curve). In this case x approaches the unitary value as v is decreasing. This is at the origin of the apparent or dynamical negative differential conductance inferable from the i-v plot. The qualitative agreement between Figs. 8-9 (blue curves) and Figs. 4-5 (red curves) support our claim according to which a form of the proposed window, particularly the one associated to matched threshold voltages with 0 value, is an approximation to Bioleks window for large p. Memristor behavior of Fig. 3(b) in [3] Let the system have an on/off conductance ratio Gon G1f = 25 and be excited by (6) with v0 = 2 V and of = 0.01 (yielding f = 2 Hz). The state initial condition is x(0) = 0.1 and vth,1 = vth,0 = vth = 0 V in (36)-(38). Under this parameter setting the BCM captures the hard-switching scenario without dynamical negative differential conductance

12

Fig. 3 a in 3 Fig. 3 b in 3 Fig. 3 c in 3

i i0 1 1.0

0.5

((t,1 ), Gon ) and Q = ((t ), Gon ) (red points) overlap and the i-v plot observed in this dynamical scenario. Recalling the discussion from Section II-C, it is obvious that Bioleks model is unable to capture these dynamics (the author himself admitted it in [12]). By contrast, Joglekars equations may replicate this behavior (see red curves in Figs. 2-3 in Section II-B). B. Modeling the i v response of other memristor nanostructures The following simulations, where we use an ordinary differ-

1.0

0.5

0.5

1.0

v v0

1 ential equation solver to integrate the BCM and, unless stated

0.5
Figure 9. Current-voltage characteristics corresponding to the memductanceux relations of Fig. 8. Blue curve is similar to red curve in Fig. 5 and to Fig. 3(a) from [3]. Black curve is similar to Fig. 3(b) from [3]. Red curve is similar to red curve in Fig. 3 and to Fig. 3(c) from [3].

reported in Fig. 3(b) of [3]. As in the previous scenario, x momentarily halts at both boundaries over period T . Conditions for single-valuedness of W = W () (i.e. (45) and (46)) do not hold. With regards to (45), this is demonstrated by the different location of P = ((t,1 ), Gon ) and Q = ((t ), Gon ) (black points) on the memductance-ux characteristic, reported in Fig. 8 (black curve). As a result, as it was observed in [12], these dynamics may not be captured by Joglekars model (see Table I). By contrast, Bioleks model may detect them. Due to the doubling of v0 and the decrease in Gon G1f of as compared to the previous scenario, here x approaches the unitary value when v is still increasing. This explains why the i-v plot of Fig. 9 (black curve) does not exhibit a dynamic negative differential conductance.

Memristor behavior of Fig. 3(c) in [3]

Our model is also capable to detect the hard-switching scenario characterized by strong nonlinearities in ionic transport and reported in Fig. 3(c) in [3]. In the following BCM simulations the input shape, its parameters v0 and , the state initial condition and the system on/off conductance ratio are set as in the simulations from Section II. However, here we take vth,0 = vth,1 = vth = 0.5 v0 V in (36)-(38). In this case x(t) temporarily stops only at boundary x = 1 over period T . Under this choice for the matched threshold voltages, letting tend set as an integer multiple of T , (47) with = 1, ti = ts and x(ts ) measured as 0.01, and (48) numerically yield t,1 = 41.7 and t = 58.3 respectively. Thus condition (45) from Section III-C is satised. Figs. 8-9 (red curves) respectively show the singlevalued memductance-ux characteristic over which P =

otherwise, we choose , D, and Gon as in Section II-A, are aimed at showing how the BCM may qualitatively reproduce other memristor behaviors recently observed in various nanoscale lms ([17], [18], [19]). In particular, the current-voltage characteristic of a metal/strongly correlated electron system (SCES) interface of Fig. 3(b) in [17] (or, alternatively, of Fig. 8(g) in [20]) is detected by our model when it features an on/off conductance ratio Gon G1f = 25 and a state initial condition x(0) = 0.6, of is driven by control source (6) with v0 = 2 V and = 0.003, i.e. f = 0.6 Hz, and the threshold voltages specifying its boundary behavior are set to vth,0 = vth,1 = vth = 0.25v0 V (see Fig. 10). Finally, simulating (1)-(2) with Gon = 2 103 S (yielding q0 = 2 105 C), Gon G1f = 5, x(0) = 0.7, usof ing window (39) with boundary conditions (36)-(38) having mismatched threshold voltages, i.e vth,0 = 0.75v0 V and vth,1 = 0.875v0 V , and exciting the system with input sinewave (6), where v0 = 4 V and = 0.01 (implying f = 4 Hz), yields plots (a) and (b) of Fig. 11 for a memristor with polarity coefcient set to 1 and +1 respectively. Plot (a) resembles the current-voltage characteristic in Fig. 2b of [19], relative to a bipolar memristor with a Pt/solide electrolyte/Cu stack proposed as basic building block for the realization of passive nano-crossbar memories. Plot (a) also replicates the currentvoltage behavior in Fig. 8(b) of [20], referring to a 300 nmthick epitaxial SrZrO3 lm doped with 0.2% Cr grown on a SrRuO3 bottom electrode with the top Au electrode extending over an area of 200 200 m2 , rst introduced for memory applications in [18]. Plot (b) of Fig. 11 resembles the current-voltage characteristic of the previously described stack [19] and thin-oxide lm [18] with reversed layer order. In each of the scenarios of Figs. 10-11, x momentarily stops at both boundaries and the memductance-ux characteristic is multi-valued, since conditions for single-valuedness (45)-(46) are not satised. C. Memristor-based circuit In this section we show agreement between simulation results on a memristor-based oscillator (see Fig. 12), derived by modeling a pair of identical memristors m1 and m2 used in the oscillator through the proposed BCM with matched zerovalued threshold voltages and Bioleks model for p = 10 respectively. The circuit is derived from Chuas canonical

13

1
1 1.5 1

anti parallel combination of memristive elements L G

0.8
0.5

iL

1 + C1 v1 0 im1 m1 im2 m2 GN2

0.6

1 v v0

+ GN1 C2 v2

0.5 0 0.5 10000

i i1 0

0.4 0.2 0 0.2 0.4 1

0.5 1 9000

9200

9400

9600

9800

Figure 12.

Memristor-based oscillator.
(a) : F , vth = 0

(b) : FB , p = 10
1.5 x 10
4

1 0.5 v v0 0

0.5

1.5

x 10

Figure 10. Current-voltage characteristic replicating the dynamics experimentally observed in a metal/SCES interface by Oka and Nagaosa (see Fig. 3(b) in [17] for the experimental result). Inset: time waveforms of x and 1 v v0 .
x2

0.5

0.5

x2
1 0.5 0 0.5 1 1.5 4 x 10

0.5

(a) : = 1

1.5

(b) : = +1

0.5

0.5

i i1 0

0.5

i i1 0

0.5

1.5 1.5

1.5 1.5

0.5

0.5

x1

x1

1.5 4 x 10

Figure 13. Projection of chaotic attractor on state plane x1 -x2 for the proposed model with vth = 0 V (plot (a)) and for Bioleks model with p = 10 (plot (b)).

0.5

1.5

0.5

1 v v0

0.5

0.5

1 v v0

0.5

inductor and the differential equation governing the rate of ionic drift for m1 and m2 : dx1 d dx2 d dx3 d dx4 d dx5 d ( ) = G1 (W (x4 ) + W (x5 ) + GN 2 ) x1 x3 , = x2 + x3 , = (x3 + x1 x2 ), = = 1 W (x4 )x1 F (x4 , 1 W (x4 )x1 , p), 0 i 2 W (x5 )x1 F (x5 , 2 W (x5 )x1 , p), 0 i

Figure 11. Current-voltage behaviors originating from the BCM with set to 1 (plot (a)) and +1 (plot(b)) respectively. Such behaviors qualitatively capture the dynamics of complementary resistive switches with memory capability ([18], [19]).

oscillator by replacing Chuas diode with the parallel between negative conductance GN 2 and the anti-parallel combination of m1 and m2 (1 = +1, 2 = 1). Let us denote the currents through m1 and m2 as im1 and im2 respectively, the voltages across capacitors C1 and C2 as v1 and v2 respectively, and the inductor current as iL . Noting that w1 D1 and (1 w2 ) D1 denote the normalized length of conductive layer for m1 and m2 respectively, dening state variables as x1 = v1 , x2 = v2 , x3 = G1 iL , x4 = w1 D1 1 and x5 = (1w2 )D1 , letting = C2 C1 , = C2 L1 G2 1 and = GN 1 G , and further considering normalized 0 time variable = t t1 , where t0 = G1 C2 is the system time scale, the following state equations may be determined after combining the constitutive equations for capacitors and

(50)

where im1 = W (x4 ) x1 and im2 = W (x5 ) x1 , F is any window function modulating the rate of ionic motion in each 0 memristor and 0 = q0 t1 . i Setting circuit element values to G = 3.3mS, GN 1 = 0.4mS, GN 2 = 1.2mS, C1 = 50nF , C2 = 37nF and L = 100mH, system parameters are numerically given by = 0.74, = 0.0333 and = 0.12. Further Gof f = 0.06mS and Gon = 1.9mS. Choosing for and D the same values specied in Section II-A and using an ordinary differential equation solver for numerical integration, chaos arises in

14

dynamical system (50) (see Fig. 13, where initial conditions for state vector x = [x1 , x2 , x3 , x4 , x5 ] R5 are set to [0.006, 0.02, 0.3, 0, 0] ). Plots 13(a) and 13(b) respectively refer to the insertion of (39) with vth,0 = vth,1 = 0 V in (36)-(38) and of (5a)-(5b) for p = 10 in place of window F in equation (50). This further proves the analogy between the BCM for matched zero threshold voltages and Bioleks model for large p. V. C ONCLUSIONS AND FUTURE RESEARCH
DEVELOPMENTS

In this manuscript the BCM, a novel, simple, accurate and general model for memristor nano-structures, was presented. With respect to the existing models, the novelty of the BCM lies in its boundary condition tuning capability, which allows one to capture both single-valued and multi-valued state-ux relations under certain sign-varying inputs. Necessary and sufcient conditions for single-valuedness of the state-ux characteristic arising from the BCM were determined for a periodic zero-mean input with one change of polarity per period. Furthermore, the BCM exhibits a closed-form solution for x(t) under any input/state initial condition combination. The BCM is capable to capture a wide class of memristor dynamics from distinct physical nano-structures, including all behaviors reported in [3]. With regards to the boundary condition tuning mechanism, for a given set of experimental data observed under a specic input/state initial condition combination, it is possible to devise an automatic procedure enabling the determination of optimized boundary conditions allowing the minimization of the mean square error between observed and modeled data. R EFERENCES
[1] L. O. Chua, Memristor: the missing circuit element, IEEE Trans. on Circuit Theory, vol. 18, no. 5, pp. 507-519, 1971. [2] L. O. Chua, and S. M. Kang, Memristive devices and systems, Proc. IEEE, vol. 64, no. 2, pp. 209-223, 1976. [3] D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, The missing memristor found, Nature, vol. 453, pp. 80-83, 2008. [4] R. S. Williams, How we found the missing memristor, IEEE Spectr., vol. 45, pp. 28-35, 2008. [5] T. Raja, and S. Mourad, Digital logic implementation in memristorbased crossbars: a tutorial, IEEE Int. Symp. Electronic Design, Test & Appl., pp. 303-309, Jan. 2010. [6] S. H. Jo, K. - H. Kim, and W. Lu, High-density crossbar arrays based on a Si memristive system, Nanoletters, vol. 9, no. 2, pp. 870-874, 2009. [7] G. S. Snider, Spike-timing-dependent learning in memristive nanodevices, IEEE/ACM Int. Symp. on Nanoscale Architectures, pp. 85-92, 2008. [8] A. A, A. Ayatollahi, and F. Raissi, Implementation of biologically plausible spiking neural network models on the memristor crossbar-based CMOS/nano circuits Proc. of IEEE Eur. Conf. Circ. Theory Design, pp. 563-566, 2009. [9] F. Corinto, A. Ascoli, M. Gilli, Nonlinear dynamics of memristor oscillators, IEEE Trans. on Circuits and Systems I, vol. 58, no. 6, pp. 1323 1336, June 2011, ISSN: 1057-7122, DOI: 10.1109/TCSI.2010.2097731 [10] Y. N. Joglekar and S. T. Wolf, The elusive memristor: properties of basic electrical circuits, Eur. J. Phys., vol. 30, pp. 661-675, 2009. [11] T. Prodromakis, B. P. Peh, C. Papavassiliou, and C. Toumazou, A versatile memristor model with nonlinear dopant kinetics, IEEE Trans. on Electron Devices vol. 58, no.9, pp. 3099-3105, Sept. 2011 [12] Z. Biolek, D. Biolek, and B. Biolkov, Spice model of memristor with a nonlinear dopant drift, Radio Eng., vol. 18, no. 2, pp. 210-214, 2009. [13] A. G. Radwan, M. A. Zidan and K. N. Salama, HP Memristor Mathematical Model for Periodic Signals and DC, IEEE International Midwest Symposium on Circuits and Systems (MWSCAS), pp. 861-864, Seattle, USA, Aug. 2010.

[14] A. G. Radwan, M. A. Zidan and K. N. Salama, On the mathematical modeling of memristors, IEEE International Conference on Microelectronics, (ICM), Cairo, Egypt, Dec. 2010. [15] W. Cai, F. Ellinger, R. Tetzlaff, and T. Schmidt, Abel dynamics of titanium dioxide memristor based on nonlinear ionic drift model, arXiv:1105.2668v1 [cond-mat.mes-hall], May 2011 [16] Y. Ho, G.M. Huang, L. Peng, Dynamical Properties and Design Analysis for Nonvolatile Memristor Memories, IEEE Trans. on Circuits and Systems I, 2011, ISSN: 1057-7122, DOI: 10.1109/TCSI.2010.2078710 [17] T. Oka and N. Nagaosa, Interfaces of correlated electron systems: proposed mechanism for colossal electroresistance, Phys. Review Letters, vol. 95, no. 26, pp. 64031-64034, 2005 [18] A. Beck, J. G. Bednorz, Ch. Gerber, C. Rossel and D. Widmer Reproducible swicthing effect in thin oxide lms for memory applications, Applied Physics Letters, vol.77, no.1, pp. 140, 2000 [19] E. Linn, R. Rosezin, C. Kgeler, and R. Waser, Complementary u resistive switches for passive nanocrossbar memories, Nature Materials, vol . 9, pp. 403-406, May 2010. [20] L. O. Chua, Resistance switching memories are memristors, Applied Physics A, vol. 102, no. 4, pp. 765-783, 2011 [21] A. D. Polyanin and M. D. Freidman, Handbook of exact solutions for ordinary differential equations, 2nd edition, Chapman & Hall, 2003 [22] F. Corinto, A. Ascoli and M. Gilli, Symmetric charge ux nonlinearity with combined inherently asymmetric memristors, Proc. of Eur. Conf. on Circuit Theory and Design, Link ping, Sweden, pp. 653-656, Augusto September 2011

Fernando Corinto (M03 SM10) received the Laurea degree in electronics and the Ph.D. degree in electronics and communications engineering in 2001 and in 2005, respectively, and the European PLACE Doctorate in 2005, all from the Politecnico di PHOTO Torino,Torino, Italy. In 2004, he won a Marie Curie HERE Fellowship (within the Marie Curie Actions under the Sixth Framework Programme). He is currently an Assistant Professor of Circuit Theory with the Dipartimento di Elettronica, Politecnico di Torino. His research activities are mainly in the areas of nonlinear dynamical circuits and systems and locally coupled nonlinear/nanoscale networks. He is coauthor of more than 80 international journal and conference papers. He has been reviewer of several papers for international journals and conferences and chair of sessions in international conferences. He is the principal investigator of several research projects. Since 2010, he is Senior Member of the IEEE, Member of the IEEE CAS Technical Committees on: Cellular Nanoscale Networks and Array Computing and Nonlinear Circuit and Systems. Fernando Corinto is the Technical Program Chair for the 13th Workshop on Cellular Nanoscale Networks and their Applications and 3rd Symposium on Memristor. He is also Visiting Professor at PPCU of Budapest, since 2007.

Alon Ascoli received the M.S. degree (First Honours) in electronic engineering from Universita degli Studi Roma Tre, Rome, Italy, in 2001 and the Ph.D. degree in electronic engineering from University PLACE Col- lege Dublin, Dublin, Ireland, in 2006. From PHOTO 2006 to 2009 he enjoyed a fruitful work experience HERE in the Swedish high-tech industry, which allowed him to get extensive expertise in state-of-the-art analog integrated circuit design and layout. Since 2009, he is with the Department of Electronic Engineering, Politecnico di Torino, Torino, Italy, where he was granted a four-year Postdoctoral Fellowship. He currently teaches undergraduate level courses on Circuit Theory and gives postgraduate level seminars on the Theory of Nonlinear Dynamics. His research interests lie in the area of nonlinear circuits and systems, networks of oscillators, cellular nonlinear/neural networks. Dr. Ascoli was honored with the International Journal of Circuit Theory and Its Applications 2007 Best Paper Award for his publication entitled Modelling the dynamics of log-domain circuits. He has served as Reviewer for the Inter- national Journal of Circuit Theory and Its Applications from 2005 to 2006.

You might also like