Modeling

3.3 Modeling

As explained before, by applying the voltage pulses $VSET$ and $VCOND$ , a (desired) high-to-low resistance switching (shown in bold in Table 3.2) is enforced in $T$ only in State 1. However, the current flowing through the memristors tends to decrease their electrical resistances and change the internal state variable $w$ . This phenomenon is called “state drift” (SD) [162] and its accumulation after a specific number of sequential (N)IMP operations causes an undesired switching event (computation error) either in $S$ or $T$ . This is due to the fact that, although the TiO $2$ memristive switches are used as two-resistance-state devices for binary data storage, they actually act as analog elements since the parameter $w$ changes continuously [142]. The design procedure of the IMP gate involves determining the proper values of the circuit parameters ( $RG$ , $VSET$ and $VCOND$ ) to minimize the SD errors (SDEs). The design procedure presented in [162], which is the only existing design procedure to the author’s knowledge, is based on a linear ionic drift model for the TiO $2$ memristive devices described below. However, as it is shown in the following, a more accurate model of the TiO $2$ memristive device has to be employed to analysis and optimize the stateful logic gates.

According to Fig. 3.2, the voltage drops on $S$ and $T$ are given by

v_S	= V _COND - V _G = i_SM_S	(3.4a)
v_T	= V _SET - V _G = i_TM_T,	(3.4b)

where

V = (i + i )R . G S T G

(3.5)

$V G$ denotes the voltage drop on $R G$ and $i S$ ( $i T$ ) and $M S$ ( $M T$ ) are the current and the memristances of the memristive devices $S$ ( $T$ ), respectively. As, in general, the memristance is a nonlinear resistance which depends on the historic profile of the current (voltage) applied to the memristor, $M S$ and $M T$ are a function of $iS$ ( $vS$ ) and $iT$ ( $vT$ ) as well as their initial resistance states (the logic input pattern). Therefore, in order to optimize the implication gate and to investigate the switching behavior of $S$ and $T$ , one has to solve Eq. 3.4 coupled with an appropriate memristor device model which accurately describes the $i - v$ characteristics of $S$ and $T$ . In the following, two TiO $2$ memristor device models and simulation studies obtained from these models are presented.

Figure 3.3.:

Schematic of the TiO $2$ memristive device cross section.

Fig. 3.3 shows a schematic of the TiO $2$ memristor structure containing a sandwiched TiO $2$ thin film and two platinum (Pt) electrodes. During an electroforming process, as dopant acting oxygen vacancies are created in the TiO $2$ thin-film except a narrow tunnel barrier of $w$ [138]. Therefore, the thin film is divided into a (high conducting) doped region and an (insulating) undoped region and its total resistance (internal resistance) is equal to the sum of the variable resistances on each region:

Rint = Rdoped + Rundoped

(3.6)

The resistance switching mechanism of the device is related to the drift motion of dopants (oxygen vacancies) due to an electric field across the device[114]. Therefore, this device shows memristive behavior as the width of the undoped region ( $w$ in Fig. 3.3) and thus the total electrical resistance of the device depends on the historic profile of the applied voltage/current to the device. According to the mathematical definition of a memristive system (Eq. 2.14), $w$ is an internal state variable which here describes the effective width of the undoped region and determines the resistance state of the memristive device.

3.3.1 Linear Ionic Drift Memristive Model

Based on Eq. 2.14 and Eq. 2.15, a simple linear ionic drift model [69] describes the internal resistance as

( ) -w-(t) w-(t)- Rint = Ro ff + 1 - Ron wmax wmax

(3.7)

and

dw-- μvRon-- = i(t), dt wmax

(3.8)

where $Ro ff$ and $Ron$ are the maximum and the minimum resistances, respectively, $wmax$ denotes the maximum thickness of the undoped region, and $μv$ is the average mobility of the oxygen vacancies in the TiO $2$ thin-film. According to the linear ionic drift model, by time-integrating the state equation (Eq. 3.8), we obtain that the modulation of the state variable $w$ is proportional to the charge passing through the device ( $Δw ∝ Δq$ ).

In order to compare the switching dynamic behavior (including the switching time and the switching energy) predicted by the linear ionic model with some recently obtained experimental data, let’s define $ΔQ$ as the amount of charge that by passing through the memristor modulates the memristance from its minimum to its maximum value ( $Δw = wmax$ ). A voltage pulse with a fixed amplitude of $v0$ and the duration of $τ0$ is applied to the memristor to perform an on-to-off (OFF) switching.

∫ τ0 ∫ τ0 ----v0--- ΔQ = i(t)dt = R (w )dt, 0 0 int

(3.9)

where according to Eq. 3.8 we have

i(t)dt = wmax---dw. μvRon

(3.10)

Therefore, $ΔQ$ is obtained as

∫ wmax 2 -wmax--- -w-max-- ΔQ = μ R dw = μ R . 0 V on V on

(3.11)

According to Eq. 3.11, derived from the linear ionic drift model, $ΔQ$ has a constant value and is independent of the $v0$ and $τ0$ . Therefore, it predicts that the switching time is inversely proportional to the voltage pulse amplitude ( $- 1 τ0 ∝ v0$ ). Indeed, according to the State-dependent Ohm’s law for a memristive system (Eq. 2.14), at a time $t0$ ( $0 < t0 < τ0$ ) where the electrical resistance of the memristor is $R0$ ( $Ron < R0 < Ro ff$ ), the rate of the charge flow ( $i$ ) is directly proportional to the voltage pulse amplitude ( $v0$ ). As a result, the amount of charge flowing through the memristor is proportional to the product of the voltage level and the time ( $Δq ∝ v0Δt$ ) and thus we have $ΔQ ∝ v0τ0$ . As according to the memristor linear ionic drift model $ΔQ$ is a constant, the switching time required for a complete switching from $Ron$ to $Ro ff$ must be inversely proportional to the pulse amplitude ( $τ0 ∝ v - 1 0$ ).

The OFF-switching energy consumption is also obtained as function of $ΔQ$ by

∫ τ0 Eswitch = v0i(t)dt = v0 ΔQ. 0

(3.12)

This predicts an inverse relationship between the switching energy and the switching time as $Eswitch ∝ v0 ∝ τ - 1 0$ . However, these predictions regarding the switching dynamic behavior ( $- 1 τ0 ∝ v 0$ and $- 1 Eswitch ∝ τ0$ ) are quite inconsistent with experimental data which demonstrate an inverse log-linear relationship between the switching time and the voltage pulse amplitude as $τ0 ∝ exp (- v0)$ [149] and a direct log-log relationship between the switching energy and the switching time as $log (Eswitch ) ∝ log (τ0)$ [75]. The main reason is that the switching dynamic behavior is significantly affected by the electron tunneling effect through the insulating region which exponentially decrease the total electrical resistance of the TiO $2$ memristor device during the switching [138]. Therefore, although the linear ionic drift model has been used to simulate the electrical properties of the memristor for different applications [129, 130, 131, 132, 133, 87, 134], in a high voltage regime, however, the tunneling effect dominates the memristor $i - v$ characteristics. For the sake of (acceptable) fast switching, the applied voltage levels in logic applications are so high that the memristive devices act as digital switches with two resistance states of $Ron$ and $Ro ff$ . As a result, it is necessary to use a more accurate model for digital (logic) applications. The switching dynamic behavior obtained from the nonlinear ionic drift model [138] explained below, shows a good agreement with experimental data. For example, this model predicts a direct log-log relationship between the switching energy and the switching time ( $log (Eswitch) ∝ log (τ0)$ ) for both micro-scale ( $5 × 5 μm2$ ) and nano-scale ( $2 50 × 50 nm$ ) TiO $2$ memristive devices (see Fig. S7 in Supplementary Information of [138]).

3.3.2 Nonlinear Ionic Drift Memristive Model

To the author’s best knowledge, the nonlinear ionic drift model presented in [138], which uses the Simmons $i - v$ expression for the insulating region as a rectangular barrier with image forces [159], is up to now the most accurate model for the TiO $2$ memristive devices. By using physically reasonable parameters, it properly describes both the static electric conduction as well as the switching dynamic behaviors and provides a good fit to the experimental data from micro-scale and nano-scale TiO $2$ memristive devices which exhibit switching behaviors effectively insensitive to the device size [138].

According to this model, the voltage across the thin-film is given by

vint = iRdoped + vg

(3.13)

where $v g$ is the voltage across the insulating region which acts as a tunneling barrier and $i$ is the current flowing through the device and its functional form is determined by [138]

[ ( √ -------)] j0A (- B √ϕI) - B ϕI+e|vg| i = ----2- ϕI e - (ϕI + e|vg |) e . Δw

(3.14)

The quantities from Eq. 3.14 are given by [138]

--e-- j0 = , Δw = w2 - w1, 2πh

(3.15)

( ) 1.2 λw 9.2 λ w1 = -------, w2 = w1 + w 1 - --------------------- , ϕ0 3ϕ0 + 4λ - 2e |vg|

(3.16)

e2ln (2 ) λ = ---------, 8 πk ε0w

(3.17)

√ ---- 4-πΔw-----2m-- B = h ,

(3.18)

( ) w1-+--w2-- 1.15-λw-- w2-(w-----w1-) ϕI = ϕ0 - e|vg| w - Δw ln w (w - w ) , 1 2

(3.19)

where $A$ describes the insulating region area, $e$ is the electron charge, $h$ is Planck’s constant, $ϕ0$ is the barrier height, $k$ is the dielectric constant, and $m$ is the electron mass. The modulation of the effective insulating region width $w$ with respect to the device current has been expressed

for $i > 0$ (OFF switching):

( ) [ ( ) ] dw i w - ao ff |i| w ----= fo ff sinh ---- exp - exp ----------- --- - --- dt ioff wc b wc

(3.20)

with the fitting parameters $foff = 3.5 ± 1 μm ∕s$ , $ioff = 115 ± 4 μA$ , $aoff = 1.20 ± 0.02 nm$ , $b = 500 ± 90 μA$ , and $wc = 107 ± 4 pm$ ;

and

for $i < 0$ (ON switching):

( ) [ ( ) ] dw i w - aon |i| w ---- = fon sinh ---- exp - exp --------- - --- - --- dt ion wc b wc

(3.21)

with the fitting parameters $fon = 40 ± 10 μm ∕s$ , $ion = 8.9 ± 0.3 μA$ , $aon = 1.80 ± 0.01 nm$ , $b = 500 ± 90 μA$ , and $wc = 107 ± 3 pm$ for physical TiO $2$ memristive devices characterized in [138]. The model fits the experimental data using the device parameters determined as $ϕ0 = 0.95 ± 0.03 eV$ , $4 2 A = 10 ± 2500 nm$ , $k = 5 ± 1$ , $Rdoped = 215 Ω$ [138].

The total resistance of the device (memristance) is equal to

vout- |vg-| Rtotal = = Rint + RPt = Rdoped + + RPt i i

(3.22)

where $vout$ is the applied voltage on the memristor, $i$ is the current flowing through the device, and $R = 2.4k Ω Pt$ accounts for the Pt electrodes resistance [139].

Figure 3.4.:

(a) $M - V$ characteristics of the TiO $2$ memristor for different values of $w$ . (b) $M - w$ characteristics plotted for a readout voltage of 0.2 V.

This model properly describes the nonlinear switching dynamics arising from the ionic motion which modulates the effective width of the insulating region (Eq. 3.20 and Eq. 3.21) as well as the electron tunneling effect through the insulating region which is a function of the width of the insulating region and the applied voltage/current (Eq. 3.14). Eq. 3.20 and Eq. 3.21 successfully model the nonlinear drift velocity of ionized dopants [138] featuring an exponential dependence on the applied current/voltage [148] and the asymmetric switching behavior [138] caused by the voltage polarity dependent competitive or cooperative behavior of ionic drift and diffusion [147].

Fig. 3.4a shows the initial memristances (before any modulation in $w$ ) for different values of $w$ as a function of the applied voltage. It illustrates that the instantaneous value of the memristance exponentially decreases with respect to the applied voltage during the switching. In fact, the electrical resistance of the device depends on the applied voltage. Therefore, low-voltage measurements are used to readout the memristance [75], which not only allows us to provide non-destructive reads but also reduces the tunneling effect. Due to the coupling between Eq. 3.14 and Eq. 3.22, the memristance needs to be calculated numerically. Fig. 3.4b shows the memristance as a function of $w$ when a readout voltage of 0.2 V is applied to the device. It illustrates that a sub-nanometer modulation in the insulating region effective width provides a $kΩ - M Ω$ readout ON-OFF-switching regime [138].

Figure 3.5.:

Memristance profile of the TiO $2$ memristive device during a high-to-low resistance switching according to the linear and nonlinear models.

Fig.3.5 shows the dependence of the memristance during ON (high-to-low resistance) switching as it follows from the linear and nonlinear ionic drift models. Due to a high voltage level applied ( $V = 1.5V$ ), the tunneling effect trough the insulating undoped region dominates the memristor $i - v$ characteristics [138]. Therefore, during the switching, the total resistance is even lower than $Ron$ , in contrast to the behavior predicted by the linear model according to Eq. 3.7. From Fig.3.5 one can see that the TiO $2$ component announced by Hewlett Packard (HP) in 2008 [69], at least at high voltage regimes, is rather a memristive system [70] than a memristor as its electrical resistance at a time is not only a function of historic profile of the applied voltage/current but also of the instantaneous value of the applied voltage/current due to electron tunneling. Furthermore, Fig. 3.5 demonstrates that when the linear model to obtain the $i - v$ characteristics of $S$ and $T$ , the voltage drops $vS$ and $vT$ (Eq. 3.4) will be wrongly predicted and the implication gate optimization will not be reliable. The linear model can be useful only in analysis and design of low voltage applications ( $- 0.5V < vM < 0.5V$ ) where the tunneling is negligible.

[next] [prev] [prev-tail] [front] [up]