Thesis Report: Exploring Memristor Topologies

Author: Serdar Benderli Adviser: Todd A. Wey School: Lafayette College

Abstract
Memristors may play a crucial role in next generation computational paradigms. By having access to such a device, circuits with new functions and new architectures of classic functions can be designed. The purpose of this thesis is to explore a number of novel topologies that make use of memristive behavior. However, due to the current unavailability of off-the-shelf memristors and lack of software simulation support for such devices, creating a behavioral model of a memristor is necessary. The first half of this thesis deals with the creation of a behavioral model. The second part uses this model to simulate and analyze three different memristor topologies, demonstrating potential of memristors and the robustness of the model. By having a behavioral model of these devices, researchers and designers can simulate circuits which involve memristors easily and efficiently, accelerating memristor research.

Page2of59

Table of Contents
Abstract ...................................................................................................................................................... 2 Acknowledgements ................................................................................................................................. 4 I. II. III. IV. V. VI. Introduction to Memristors.......................................................................................................... 5 Thin-Film TiO2 Memristors ........................................................................................................... 8 Potential Applications of Memristors ..................................................................................... 10 Current Problems in Memristor Research............................................................................. 11 Behavioral Modeling of Thin-Film TiO2 Memristors ........................................................... 13 Topologies ..................................................................................................................................... 23 1. Memristor Switch ................................................................................................................ 24 2. Memristor AM Modulator ................................................................................................... 32 3. Q-Factor Controller for 2ND Order Band-Pass Filter ...................................................... 49

VII. Conclusions................................................................................................................................... 57 Bibliography ............................................................................................................................................ 59

Page3of59

Acknowledgements
I would like to acknowledge and extend my heartfelt gratitude to the following persons who have helped me complete this thesis: My adviser, Prof. Todd A. Wey, for all his hard work, encouragement, and support throughout the year. He was always there to help during the countless problems I ran into. He has endured many hours trying to fix my silly syntax errors and he never yelled. His enthusiasm has only gone up throughout the year, and this has always kept me wanting to deliver more than I otherwise would have. Without him, this thesis would not be possible. Prof. William D. Jemison, for all his important input throughout the year, for accepting to be a part of my thesis committee, and for his comments and corrections on the final draft of this thesis. Prof. Jeffrey O. Pfaffmann, for accepting to be a part of my thesis committee and for his valuable input on the final draft of this thesis All Electrical and Computer Engineering faculty members and Staff and to my family and Raluca for their constant support and encouragement.

Page4of59

I.

Introduction to Memristors

The term Memristor was first coined by Prof. Leon Chua in a 1971 paper [1] as a short for memory resistor. In traditional circuit theory, the resistor provides the relationship between current and voltage, the capacitor provides the relationship between voltage and charge, the inductor provides the relationship between current and flux, i =

dq , and dt

v=

d , but no relationship exists between flux and charge. In order to have a dt

conceptual symmetry of relationships between four fundamental electrical properties, Chua postulated that a fourth circuit element should exist which would provide this missing relationship.

Figure 1 - Fundamental Electrical Properties and Defining Relationships ImageCourtesyof:J.J.Yang/HPLabs

Page5of59

Her argued that if flux is a function of charge, = F (q) , then when both sides of this equation are differentiated, we get:

d ( q ) F ( q ) dq = dt dt dt

(1)

where

F (q) can be re-written as a function of q(t), M(q(t)). Chua defined this as the dt

memristance. If we replace the derivative of flux with voltage, and the derivative of charge with current, we get a more familiar equation:

v (t ) = M (q (t ) )i (t )

(2)

As can be seen, memristance has the same units as resistance, ohms. It is important to note that the memristor behaves like a resistor at any given time t. However, its resistance depends on the history of the current that has passed through it. The basic function = F (q) can be re-written in the inverse form, q = F 1 ( ) given that it is invertible. Then, by taking the same steps as before, we can arrive at the equation:

i (t ) = W ( q (t ))v (t ) , where W (q(t )) can be defined as the memductance.

After laying out the basics, Chua moves on to talk about active-circuits he designed by using transistors and other elements, to behave like a memristor. However, no one was able to find a physical manifestation of a memristor to serve as a passive device for over 30 years, until HP Labs did in 2008. They reported their findings in a paper published in Nature [2]. After HP Labs findings, there have been numerous other reports of devices that show memristor-like behaviors. Since it is out of the scope of this
Page6of59

paper, they will not be talked about in detail. Please refer to the bibliography for a list of such devices [3, 4, and 5]. The author has chosen to work with TiO2 memristors because they are better documented than the other reported realizations are and they show potential of being available commercially in the near future.

Page7of59

II.

Thin-Film TiO2 Memristors

While working on thin-film TiO2, HP Labs researchers noticed hysteresis in the i-v curves [2]. This implied that the resistance of the thin-film device depended on the history of current that had passed through it, essentially resulting in memory.

Figure 2 An Array of 17 TiO2 Memristors (Left) and 2-dimensional representation of a thin-film TiO2 device (Right) Images Courtesy of: HP Labs

The memristive behavior of thin-film TiO2 is a direct manifestation of the shift in the boundary between oxygen-rich (undoped) and oxygen-depleted (doped) regions when an electrical field is applied [2]. The depleted region acts as if it was doped with +2 charge carriers and thus has a much lower resistance than the undoped region. When an electrical field is applied, the carriers drift with mobility (v), changing the widths (w)
Page8of59

of the low-resistance (RON) and high-resistance (ROFF) layers, effectively changing the materials overall resistance. Moreover, this drift is bounded by the length of the material (D). The equations governing the width of and the voltage across a TiO2 memristor are as presented in [2]:
R dw(t ) = v ON i (t ) dt D RON q (t ) D

(3)

w(t ) = v

(4)

w(t ) w(t ) + ROFF (1 vm (t ) = RON ) im (t ) D D

(5)

As can be seen, the memristance of the device is simply the weighted average of the two regions resistances, where the weights are the proportion of the region widths to the overall device length D. A very important property of these devices can be seen from (3): when there is no current passing through, dw(t) is 0. This means that the device retains its memory even when the power is turned off. It is also important to note that this device does not provide a direct link between flux and charge. It is simply changing its resistance as a reaction to the current passing through it.1

1 This was the source of some criticism to the HP findings, noting that it may not be correct to call such a device
memristor since it is not exactly the same as Chuas memristor [3, 4].

Page9of59

III.

Potential Applications of Memristors

The first commercial application for memristors will most likely be in the digital domain where they are used as switches. This is also the first topology that will be considered in this thesis. HP Labs predicts that such devices could reach the markets within the next five years. They already reported building a hybrid chip that contains both transistors and memristors using current manufacturing technology [6]. However, it is likely that memristors will offer their most important advantages in the analog domain when used as non-volatile analog memory units. For the first time, researchers and designers would have simple access to non-volatile analog memory which opens up numerous possibilities for new computational paradigms in the fields of signal processing, control systems, neural modeling, and neural networks. Very simple learning circuits that involve a simple RLC circuit and a memristor that model the learning behavior in amoeba are already being studied [7]. Especially in the emerging field of neuromorphic engineering, memristors may be used to create simpler and more efficient neuron models than the currently used transistor architectures [8]. More applications of memristors are likely to emerge as more people learn to design with them. However, there are still a number of other problems that impede memristor research.

Page10of59

IV.

Current Problems in Memristor Research

The biggest problem in memristor research is that there are no off-the-shelf memristor parts to build physical circuits. This problem will likely stay unresolved for at least a few more years. Until then, researchers and designers who do not have access to nanoscale manufacturing facilities must rely on a theoretical approach. There are two ways one can approach theoretical memristor research: in a purely mathematical way where memristors equations are used along with Kirchhoffs current and voltage laws to solve the circuits, or by using circuit simulators such as SPICE. The mathematical method is not widely used because even though such models may be sufficient for simple circuits, they cannot be used to solve complex circuits easily and efficiently. Thus, for complex solvers, solvers such as SPICE have become an industry standard in designing, simulating, and verifying circuits. Most circuit solvers have built-in mathematical models for basic circuit elements (primitives) such as resistors, diodes, and transistors. Furthermore, behavioral models for more complex elements, such as op-amps, can be created by using these primitives. Behavioral models are desirable since they capture the behavioral essence of devices rather than providing an exact mathematical description. Once a behavioral model of a device is made, it can be used in any circuit topology and thus makes simulation of different circuits very easy and efficient. However, no such model for memristors existed at the time the author started his research. Thus, it was necessary to build a behavioral model for memristors in order to study potentially useful memristor topologies. The model should be of great relevance to anyone who decides to study memristors using circuit simulators because it enables
Page11of59

them to easily and efficiently design, simulate, and verify any circuit design that involves memristors without having to construct and solve complicated mathematical equations.

Page12of59

V.

Behavioral Modeling of Thin-Film TiO2 Memristors2

The first thing to note about thin-film TiO2 is that equations (3), (4) and (5) are only valid when w is within the boundaries of the device. Once w reaches either end of the device, it can move no further, and the overall memristance of the device must remain constant (clipped) until current in the opposite direction is applied. This dynamic is difficult to model behaviorally. Thus, the modeling process should be broken into two stages for simplicity. First, an unconstrained model is built where boundary conditions are not imposed, and then it is expanded to include the boundary conditions. The unconstrained model is based on the following memristor relationships:

R dw(t ) = v ON i (t ) dt D
RON q (t ) D

(3)

w(t ) = v

(4)

w(t ) w(t ) vm (t ) = RON + ROFF (1 ) im (t ) D D

(5)

SPICE language was used to create the behavioral model in this paper. However, other than a few

language-specific rules, the behavioral basics of the model are language-independent and can be implemented in any circuit simulator.

Page13of59

Figure 3 shows the proposed 2-terminal model for unconstrained w, that incorporates these relationships3:

1 +

i(t)

w(t)

ROFF v(t) F(w,i(t)) Iw CW gmin

2
IWmultipliesi(t)by v RON .
D

Thisforms dw(t )
dt
RON ROFF . D

F(w,i(t))multipliesi(t)by w(t )

A1FcapintegratesIw, resultinginw(t)as voltageacrossCW.

VoltageacrossFandROFFgivesvoltage across memristor.

Figure 3 Unconstrained Model

The values for the parameters v, D, RON, and ROFF were taken directly from [2] and are 10

-10

cm2s-1V-1,

10nm, 100, and 16K respectively and are the same in all circuits and figures throughout the paper.

Page14of59

From (3),

R dw is formed by multiplying the current i(t) by v ON , which is the gain of the dt D

current-controlled-current-source Iw. To get (4), both sides of (3) need to be integrated. This is achieved by a 1F capacitor that integrates a current value of dw to give a voltage dt

of w(t) across4. This capacitance voltage value is then passed to a voltage-controlledvoltage-source F(w,i(t)) where it is multiplied with the current i(t) and a partial memristance to get (5). This partial memristance does not include the ROFF term in order to satisfy SPICE voltage loop criteria when the memristor is driven by a voltage source. The partial memristance then is given by
RON ROFF and the overall voltage across the D

memristor is simply the sum of the voltage across ROFF and the voltage across the dependent source. This completes the unconstrained circuit model and it fully describes a TiO2 memristor when w(t) is not constrained. In order to impose the boundary conditions, w(t) must be clipped when it tries to go over D or below 0. Figure 4 shows the proposed clipping circuit that is used to expand the unconstrained model described above:

A large resistor gmin is placed in parallel with the capacitor in order to satisfy SPICE convergence criteria

Page15of59

Figure 4 - Proposed Clipping Circuit

Four comparators are used in order to ensure that w(t) does not go beyond its limits. These comparators are ideal and are created by using voltage controlled voltage sources. Two pairs of comparators clip w(t) at its top and bottom boundaries. When the external bias is positive and w(t) is equal to the upper boundary, a switch is closed that connects the CW of the unconstrained circuit to a voltage source equal to D, thereby effectively clipping w(t) at D. Conversely, when the voltage source is negative and w(t) is equal to 0, a switch is closed that connects the CW to ground, thereby effectively clipping w(t) at 0. The clips are always released when the current bias polarity changes in order to allow w(t) to move away from the boundaries.

Page16of59

The following results can be compared to those reported by HP Labs in [2] to verify that the circuit models the behavior of the reported thin-film TiO2 memristors correctly. Figure 5 shows voltage and current vs. time and the hysteretic characteristic of the i-v curve when w(t) is within boundaries. Another important characteristic of these devices can be seen when the frequency of the input bias is increased. Due to the low mobility of the positive carriers, the change in w(t) is slow and cannot keep up with increasing bias frequencies. When the frequency is high enough, the change in w(t) will be unobservable, essentially biasing the memristor at the current memristance and practically turning it into a regular resistor. This is reported in [2] with the collapse of the Lissajous i-v curve, which can also be seen in Figure 5 for a tenfold increase in the frequency of the input bias from f0 to 10f0.

vm(t) im(t) w(t)/D i-v at f0 i-v at 10f0

Figure 5 - Voltage & Current vs. Time and i-v characteristics for w within boundaries

Page17of59

Figure 6 shows the results in the case where w(t) is clipped:

Figure 6 - Voltage & w(t)/D vs. Time and i-v characteristics when w is clipped

As seen in Figure 6, w(t) is clipped as soon as it tries to go beyond 0 or D and is kept there until the polarity of the input bias voltage is reversed. Further nonlinear effects that are observed in the actual devices, such as non-linear drift mentioned in [2], can also be added to the model. We modeled this nonlinear drift in two different ways. The first is to replicate a window function such as the one mentioned in [2] via additional polynomial dependent sources. The second is to increase the capacitance Cw near the boundaries via additional switches and capacitors thereby modeling the nonlinear behavior at the boundaries as a piecewise nonlinear capacitance.

Page18of59

Figure 7 shows the extra capacitance circuit:

Figure7NonlinearDriftviaAddedCapacitance

When w(t) is above a chosen value wHIGH and below the a chosen value wLOW, capacitances C1 and C2 are added in parallel with CW. When w(t) is within these bounds, the caps are kept at the voltages wHIGH and wLOW for preserve continuity when extra capacitances are added.

Page19of59

The i-v characteristics in both cases are qualitatively similar. However, the windowing method is preferable for cases where a simple windowing function can be used to model the nonlinearity. For TiO2 memristors, a simple window function can approximate the observed non-linearities as mentioned in [2]. We chose the following window function that is comparable to the one presented in [2]: w(t )(D w(t ) ) D2

(6)

Imposing this windowing on (3), we get a new equation that governs change in w: dw(t ) w(t )(D w(t ) ) RON =4 v i (t ) dt D D2

(7)

The factor of four is added in order to scale the maximum value of mobility, V, which occurs at w=0.5D, to its maximum value when the windowing is not included. This can be seen in Figure 8 where the dashed line is the mobility without the factor of four in the window function and the continuous line is the mobility with the factor of four:

Page20of59

0.9

0.8

0.7

mobility (uv)

0.6

0.5

0.4

0.3

0.2

0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9 x 10

1
-8

w (nm)

Figure8WindowingRescaling

Figure 9-10 shows the resulting i-v characteristics of both the approaches:

Figure 9 - Voltage & w(t)/D vs. Time and i-v characteristics when a windowing function is used

Page21of59


Figure10Voltage&w(t)/Dvs.Timeandivcharacteristicswhenextracapacitorsused

From these results, we can conclude that the model shows excellent agreement with the results presented in [2]. The robustness of this model will be further reinforced in the following section as the memristor will be used in three different topologies. Furthermore, the work in this section provides a behavioral modeling framework for similar memristive devices.

Page22of59

VI.

Topologies

Three topologies will be considered in this section. The first topology is the memristor switch, where the memristor is used as an on/off switch. This topology is mainly useful when the memristor is used in a NVRAM application. One of the key drawbacks of such NVRAM devices would be the slow switching speed of the memristors due to the low speed of the carriers in the TiO2 memristors. This problem will be analyzed further and comparisons will be made to current transistor switch technology to see the extent of this drawback. The second topology is the Memristor AM Modulator. In this topology, the memristor is used in a simple inverting op-amp configuration. A low frequency signal modulates a carrier signal by changing the memristance, thereby changing the overall gain of the opamp. This effectively modulates the carrier signal. Further analysis will be made in this section to evaluate the performance of this topology. The third topology is the Q-factor Controller for a 2ND order filter. In this topology, the memristors initial value of w is used to set the Q-factor of a 2ND order Band-Pass filter.

Page23of59

1. Memristor Switch
In this topology, the memristor is used alone with an independent current source as shown in Figure 11:

Figure11MemristorSwitchingCircuit

The current source is used to vary the memristance between its maximum and the minimum values as fast as possible thereby switching the memristor on and off. Since this speed depends on the size of the current, it is constrained by the material properties of thin-film TiO2. Unfortunately, no such data on the materials current limits is currently available. Thus, an exact limit on the speed is not provided in this section. However, an analysis of the size of the current for a desired speed is provided. Due to the effects of windowing near the boundaries, this topology should be analyzed separately for the windowed and the non-windowed memristors. We start the analysis with the non-windowed case since it is simpler. Suppose we wanted the memristor to switch in 10usec. From
R dw(t ) = v ON i (t ) , we can see that for a dw(t) of D, RON=100 dt D

and D=10nm, we would need a current of i(t)=10A, to get a dt, which we can call the
Page24of59

switching time ts to be 10usec. ts can also be written in terms of initial w, w0, and final w, wf, for constant current IM:

w(t ) = w0 + V

w f w0 RON I M t ts = R D V ON I M D

It can be seen that for the values we chose for w0=0, and wf=D, this equation also results in ts=10usec. A more interesting result can be seen in the power consumption which can be found by the formula P=IM2ROFF. For a constant IM=10A and ROFF=16K.we find that the peak power consumption by the device is 1.6MW,. This is indeed a huge amount, and it is proportional to the square of the current. We can integrate the P=IM2M(t) to get the total energy required for a switching with constant current:
ts ts w(t ) w(t ) ( R ROFF ) + ROFF (1 Es = I M 2 RON ) dt = I M 2 ROFF ON w(t ) dt D D D 0 0 ts

= IM = IM

R
0

OFF

R ( RON ROFF ) ( w 0 + V ON I M t ) dt D D
w f w0 ts w w0 R ) V ON I M = f ( w0t + t 2 0 D ts s
ts

RON ROFF ROFF t + D

R ROFF w f + w0 w f w0 Es = I M ROFF + ON D D 2 V RON

(8)

For the chosen values of wf=D and w0=0, this equation can be reduced to a simpler form:
D 2 1 ROFF + 1 V 2 RON

Es = I M

(9)

Page25of59

For the chosen values of ROFF and RON this equation gives us 8.05J as the energy required for a single switching of the memristor. We can manipulate (9) to write the energy in terms of ts rather than IM:

D2 D2 1 1 D 2 1 ROFF Es t s = + 1 Es (ts ) = V 2 RON V RON V 2 RON

ROFF 1 + 1 RON ts

This equation can be used to plot ES against ts to show graphically the energy required for a desired switching time:

10

10

Energy (Joules)

10

10

10

X: 1e-005 Y: 8.05

10 -9 10

10

-8

10

-7

10

-6

10

-5

ts (sec)

Figure12Energyvs.SwitchingTimeforNonwindowedMemristor

Page26of59

We can also find the average power required for a switching the memristor in ts seconds. PAVG can be defined as PAVG=Es/ts. For ts=10usec, PAVG= 805KW. Figure 13 shows the SPICE simulation with a 10A current passing through a nonwindowed memristor for 10usec:
2.0MW

POWER MEMRISTANCE

1.0MW 0W 20K 10K 0 10A

CURRENT

5A 0A 0s

0.5ms

1.0ms

TIME

Figure13SPICESimulationShowingaNonWindowedMemristorSwitching

The memristor in this case switches in 10usec as predicted. At this rate then, the memristor can be switched with a frequency of 100KHz. For a current of 100A, the memristor switches in 10usec, 1000KHz. When the power curve from Figure 13 is put into MATLAB and integrated analytically, we get the predicted energy requirement, 8.049J.

Page27of59

The windowed case is slightly more complicated to analyze mathematically due to the resulting integrals. Rearranging (7), we get the following for constant current IM:

dw(t ) w(t )(D w(t ) ) RON =4 v i (t ) dt D D2

R dw(t ) = v ON I M dt w(t )(D w(t ) ) D 4 2 D
Rearranging and integrating both sides gives us:

RON D w ln w + D = K 0 + v D I M t 4
where K0 is an integration constant resulting from the initial value of w0. Rearranging this, we get:
R 4 K 0 + V ON I M t D D

w(t ) = ( w + D)e
Finally, we can write:

w(t ) =

' DK 0 ' K 0 + e At

(10)

where A = 4V

RON ' I M and K 0 is another constant resulting from the exponentiation. By D2

' setting t=0, we can find K 0 in terms of w(0):

w(0) = w0 =

' DK 0 w0 ' K0 = ' D w0 K0 +1

Page28of59

We can then use (10) to find tsw:

' DK 0 wf = ' t sw K 0 + e At

' D wf ln K 0 wf = A

It is important to note that as we choose wf to be closer to D, tsw goes to infinity. This makes sense since if w is at the boundary, then dw/dt is 0, regardless of the current passing through. We cant choose wf to be equal to D, but we can choose it to be really close. For wf=0.999D, the above equation gives us tsw=34.53usec. We can integrate the P=IM2M(t) to get the total energy required for a switching for constant current: From (10), we can also find the energy Esw as a function of tsw:
' ' K0 K0 w(t ) w(t ) 2 + ROFF (1 + ROFF (1 ' ) dt = I M RON ' ) dt RON At At D D K0 + e K0 + e

Esw = I M

1 ' Esw = I M 2 ( RON ROFF ) t ln ( K 0 + e At ) + ( ROFF t ) A

(11)

For the chosen values of ROFF and RON this equation gives us 27.8J as the energy required for a single switching of the memristor. We can evaluate the integral in (11) from 0 to tsw and rearrange it to write the energy in terms of tsw rather than IM:

Page29of59

 D wf ln K 0' wf A = 4 RON I IM = v M R D2 4v ON tsw D2 1 1 ' ' Esw (tsw ) = I M 2 ( RON ROFF ) tsw ln ( K 0 + e Atsw ) + ln ( K 0 + 1) + ( ROFF tsw ) A A
This equation can be used to plot Esw against tsw to show graphically the energy required for a desired switching time:

10

10

10

Energy (Joules)

10

10

X: 3.346e-005 Y: 28.69

10

10 -9 10

10

-8

10

-7

10

-6

10

-5

10

-4

tsw (sec)

Figure14Energyvs.switchingtimetswforWindowedMemristor

We can also find the average power required for a switching the memristor in ts seconds. PAVGw can be defined as PAVGw=Esw/tsw. For tsw=34.53usec, PAVGw= 805KW.

Page30of59

Figure 15 shows a windowed memristor switching on and off consecutively with a current of 10A with a w0 of 0.001D:
2.0MW

POWER MEMRISTANCE

1.0MW 0W 20K 10K 0 10A

CURRENT

5A 0A 0s

0.5ms

1.0ms

TIME

Figure 15 SPICE Simulation showing a Windowed Memristor Switching

The off value of memristance is 15.984K whereas the on value is 100.8. The memristor switches on in 34.53usec. We can see the dramatic increase in the amount of time it takes for the memristor to switch when non-linear drift effects are taken into account. Peak power consumption is 15.98MW and total energy required for a single switch is 27.61J. These values are dependent on the value of w0 we choose, the farther away from the boundaries we pick the initial value, the less energy will be required to switch it. We can conclude that the model is robust and performs as expected in this topology. However, these values suggest that such an application would be unfeasible if indeed the switch is made from a single memristor where its memristance is varied by a large current source. Perhaps as HP Labs develops their memory chips in the coming years, more light will be shed on this matter, along with more data for further analysis.
Page31of59

2. Memristor AM Modulator

In this topology, the memristor is used as the feedback resistor in an inverting gain opamp configuration as shown in Figure 16. The purpose of this topology is to modulate a high-frequency carrier signal vRF, referred to as the carrier, by varying the memristance which effectively changes the gain of the inverting op-amp. The memristance is varied by a low-frequency signal vAM, referred to as the signal, which we wish to transmit. The modulation is achieved by controlling the amplitude of the carrier by changing gain of the op-amp. As with a typical AM modulator, the carrier is enveloped by the signal and can be retrieved on the receiver side back by envelope detection. The output of the opamp is the combination of the modulated carrier and the signal itself. To retrieve only the modulated carrier that we wish to transmit to the receiver, a high-pass filter is used. Due to the inherent integration within the memristor, the signal is passed through a differentiator in order to keep amplitude proportional over the circuit.

Figure 16 - Proposed Memristor AM Modulator

Page32of59

For this application to work, it is important that the signal be a sufficiently low frequency and the carrier is sufficiently high frequency. The reason for this constraint lies in the slow frequency response of the w of the TiO2 memristor due to low carrier speeds. As mentioned before, the memristor acts like a resistor when the frequency of the applied signal is high enough because the w(t) cannot keep up. Thus, the effect of the carrier on the memristance is negligible given that the low frequency signal is of sufficiently high amplitude and sufficiently low frequency so that it can change the memristance. The modulation depth depends on how much the gain of the op-amp varies overall. It is insightful to start with a mathematical derivation of what the output should be so that we can compare and justify the behavior of the model. The current through the memristor can be written as:
i (t ) = 1 d v (t ) vRF ( t ) AM R1 dt

(12)

Plugging (12) into (3) and integrating both sides, we get the following expression for w(t):

R dw(t ) = v ON i (t ) dt D

w(t ) = w(0) +

V RON
R1 D

v (
t 0

RF

(t ) dt v AM (t )

(13)

Page33of59

Using this expression for w(t), we can re-write (5) as follows:

d v (t ) w(0) 1 vM (t ) = ROFF + ( RON ROFF ) vRF ( t ) AM + D dt R1 t V RON d v (t ) 1 ( RON ROFF ) vRF (t ) dt v AM (t ) vRF ( t ) AM 2 0 R 1 R1 D dt

(14)

We also know that vout(t) is:

vopamp (t ) =

d v AM (t ) vM (t ) dt

(15)

Using these equations, we can mathematically solve for w(t) and the op-amp output voltage. For this analysis, we pick the carrier to be a 1 KHz cosine wave with amplitude 1V, R1 to be 5K, and the signal to be a 1Hz sine wave with amplitude 0.2V. Figure 17 shows the MATLAB analysis results. The top graph is the w(t) for the memristor, and the bottom graph is the vout(t) of the op-amp prior to the filtering which is the combination of the signal and the modulated carrier.

Page34of59

1 Length (m)

x 10

-8

w(t)
X: 0.7503 Y: 9.003e-009

0.5
X: 0.2377 Y: 1.009e-009

0 0

0.5

1.5

2.5

3.5

4.5

vout(t)
10

Amplitude (volts)

5 0 -5 -10

X: 0.0955 Y: 5.78

0.5

1.5

2.5

3.5

4.5

vfiltered(t)
4

Amplitude (volts)

2 0 -2 -4

X: 0.2575 Y: 2.881 X: 0.7505 Y: 0.338

0.5

1.5

2.5

3.5

4.5

Time (sec)

Figure17MATLABAnalysisResults

The peak values for w(t) are 9nm and 1nm. If vout(t) is passed through a perfect filter which eliminates the low-frequency signal component, we get the filtered vfiltered(t) shown in the bottom graph. The crest peak value of the modulated carrier is 2.88V whereas the trough is 0.32V. The modulation index can be found by m = this case. The results of the SPICE simulation for the corresponding circuit are given in Figures 18-20:

vC vT which equals 88% in vC

Page35of59


Figure18SPICESimulationofw(t)withNonWindowedMemristor

Figure19SPICESimulationofvout(t)withNonWindowedMemristor

Page36of59


Figure20SPICESimulationofvfiltered(t)withNonWindowedMemristor

As can be seen, the memristance is successfully varied at a rate of 1Hz, and the carrier is successfully modulated. Furthermore, we can see that the effect of the carrier on the memristance is negligible for the chosen frequencies and amplitudes of the carrier and the signal. The magnitude of vout(t) is as predicted by the MATLAB analysis. The slight difference between the predicted maximum peak and the maximum peak of simulated vfiltered(t) is a side-effect of the non-perfect filtering Thus, we can conclude that the model behaves according to theory.

Page37of59

The next step is to look at the performance of this circuit. This can be done by calculating the Total Harmonic Distortion (THD)5 in the signal after it has been demodulated. Two main factors contribute to THD: i. ii. Imperfect components (i.e. op-amp, filter, etc.) Modulation Depth

Since we used a non-windowed memristor in this first analysis, we would expect the best-case THD to be very close to 0. This is because for a perfectly filtered vout(t), the envelope of the filtered output is simply w(t), and there is the only distortion in w(t) is due to the high-frequency signal. Figure 21 shows the FFT of the w(t):

Magnitude Relative to Fundamental Frequency (dB)

X: 1 Y: 0

-100

X: 1000 Y: -147.5

-200

-300

-400

-500

-600 0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

Frequency (Hz)

Figure21MATLABResultofFFTontheEnvelope YaxisisMagnitudesrelativetofundamentalfrequencyindB

For a modulation index of 88%, the best-case THD is 6.34e-5%, which is very small as we expected. We can also plot THD over a range of modulation indices:

TheTHDiscalculatedtobethesumofthesquaresoftheamplitudesoftheharmonicfrequenciesoverthe squareoftheamplitudeofthefundamentalfrequency,inthiscase,thefrequencyofthesignal,1Hz.
5

Page38of59

10

10

10

-1

THD (%)

10

-2

10

-3

10

-4

X: 88.18 Y: 6.338e-005

10

-5

10

20

30

40

50

60

70

80

90

100

Modulation Index (%)

To calculate THD on the SPICE simulation, the output of the op-amp in the SPICE simulation is passed through an imperfect high-pass filter. The 1st order high-pass filter is a simple RC filter with a cutoff frequency (fc) of 100Hz. This output of the filter is then sent into MATLAB where it is passed through a perfect envelope detector. The detected envelope is the received signal that we wanted to transmit. Figures 22-23 shows the MATLAB output of the envelope detection and the FFT of the envelope. The THD for this case is 0.018%. We can conclude that the extra distortion is due to the imperfect filtering.

Figure22MATLABResultofEnvelopeDetection

Page39of59

0
X: 1 Y: 0

-100

Magnitude Relative to Fundamental Frequency (dB)

X: 2 Y: -86.79

-200

-300

-400

-500

-600

-700 0

50

100

150

200

250

300

350

400

450

500

Frerquency (Hz)

Figure23MATLABResultofFFToftheEnvelope YaxisisMagnitudesrelativetofundamentalfrequencyindB

A valid question is to ask how the non-linear windowing on the memristor affects the performance. We would expect the THD to be higher in this case due to the nonlinearity we impose on w(t). To answer this question mathematically, we must re-write equations (13) and (14) to include the windowing. The other equations from the nonwindowed analysis are still the same. In order to re-write (13), we combine (7) and (12) and integrate both sides:

w (t ) ( D ww (t ) ) RON dww (t ) =4 w i (t ) v dt D2 D
i (t ) = 1 d v (t ) vRF ( t ) AM R1 dt

(7)

(12)

Page40of59

 ww RON t D v RF (t )dt v AM (t ) = K 0 + v ln 4 ww + D R1 D 0
The subscript w shows that we are solving for the windowed case. We can now rearrange and solve for ww(t)
t RON vRF ( t ) dt v AM ( t ) 2 R1 D 0

4 v

ww (t ) =

DK e
' 1 + K 0e

' 0

4 v

t RON vRF ( t ) dt v AM ( t ) 2 R1 D 0

(16)

' where K 0 is

w0 ,evaluated the same way as in the memristor switch analysis. D w0

We can then plug (16) into (5) to get vMW(t).

Page41of59

Using these equations, we can analyze how the AM Modulator should behave for the input signals we chose. We need to make sure that the modulation index for this case is also 88%, so we can have a valid comparison between the windowed and nonwindowed cases. To achieve this, we need to use a trial-and-error approach to pick the correct amplitude of vAM(t) that will range the ww(t) between 0.9mm and 0.1mm. We would expect this amplitude to be higher than the 0.2V we used for the non-windowed case, due to the slowing carrier speeds in the windowed case. An amplitude of 0.275V gives us the desired variance of ww(t) and the below figures show the analysis of the AM Modulator for this analysis:

x 10

-8

w(t)
X: 0.7504 Y: 9.002e-009

Length (m)

0.5
X: 0.2389 Y: 1.002e-009

0 0

0.5

1.5

2.5

3.5

4.5

vout(t)
Amplitude (volts)
10 5 0 -5 -10 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
X: 0.0825 Y: 7.503

vfiltered(t)
4

Amplitude (volts)

2 0 -2 -4 0

X: 0.2615 Y: 2.882 X: 0.7505 Y: 0.3372

0.5

1.5

2.5

3.5

4.5

Time (sec)

Figure24MATLABAnalysisResultsforWindowedAMModulator

We know that the best-case THD can be found by looking at the FFT of the envelope, which is w(t) for a perfectly filtered vout(t). For a modulation index of 88%, the best case THD is 0.54%. In other words, if we used perfect components, the best THD we could achieve is 0.54%. Figure 25 shows the FFT of w(t):
Page42of59

0
X: 1 Y: 0 X: 3 Y: -51.11

-50

Magnitude Relative to Fundamental Frequency (dB)

X: 5 Y: -97.7

-100

-150

-200

-250

-300

-350

-400 0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

Frequency (Hz)

Figure25MATLABResultofFFTofww(t) YaxisisMagnitudesrelativetofundamentalfrequencyindB

We can also plot the best-case THD values over a range of modulation indices:

10

X: 0.9169 Y: 2.454

10

0 X: 88.01 Y: 0.5419

THD (%)

10

-1

10

-2

X: 29.76 Y: 0.001897

10

-3

10

20

30

40

50

60

70

80

90

100

Modulation Index (%)

Figure26THDoveraRangeofModulationIndicesforWindowedMemristor

Page43of59

As we can see, we would achieve the lowest best-case THD of 1.89e-3% for a modulation index of %29.76. As modulation index goes down, the amplitude of the highfrequency carrier rises relative to the amplitude of the low-frequency signal and thus its effects on w(t) become more significant, increasing THD. At a modulation index of 0.91%, the best-case THD is 2.45%. As modulation index goes up, THD caused by windowing goes up since w(t) gets closer to the boundaries. At a modulation index of 88%, the best-case THD is 0.54%. To see the results of the SPICE simulation, we can simply plug in the windowed memristor model into the circuit. In order to make a valid comparison between the two circuits, we must make sure that we are comparing two scenarios where the modulation indices are the same. So, we need to set the amplitude of the signal to 0.25V to achieve a modulation index of %88, as found in the previous section.

Figure27SPICESimulationofw(t)withWindowedMemristor

Page44of59


Figure28SPICESimulationofvout(t)withWindowedMemristor

Figure29SPICESimulationofvfiltered(t)withWindowedMemristor

Page45of59

As can be seen, the memristance is successfully varied at a rate of 1Hz, however the non-linear effects are apparent. The magnitude of vout(t) is as predicted by the MATLAB analysis. The slight difference between the predicted maximum peak and the maximum peak of simulated vfiltered(t) is a side-effect of the non-perfect filtering. To calculate THD on the SPICE simulation, the output of the op-amp in the SPICE simulation is passed through an imperfect high-pass filter. The 1st order high-pass filter is a simple RC filter with a cutoff frequency (fc) of 100Hz. This output of the filter is then sent into MATLAB where it is passed through a perfect envelope detector. The detected envelope is the received signal that we wanted to transmit. Figures 30-31 shows the MATLAB output of the envelope detection and the FFT of the envelope:

Figure30MATLABResultofEnvelopeDetectionforWindowedMemristor

Page46of59

0
X: 1 Y: 0 X: 3 Y: -50.47

-100

Magnitude Relative to Fundamental Frequency (dB)

X: 5 Y: -96.74

-200

-300

-400

-500

-600

-700 0

50

100

150

200

250

300

350

400

450

500

Frequency (Hz)

Figure31MATLABResultofFFToftheEnvelope YaxisisMagnitudesrelativetofundamentalfrequencyindB

The THD in this case is 0.69%, slightly higher than the predicted value of 0.54% due to imperfect filtering. We can repeat this process for the two other data points in Figure 26 to show that the circuit shows the THD characteristics predicted mathematically. A total of three points is shown in Figure 32, superimposed on the mathematically predicted curve:

Page47of59

10

10

THD (%)

10

-1

10

-2

10

-3

10

20

30

40

50

60

70

80

90

100

Modulation Index (%)

Figure32SPICESimulationResultsof3THDvaluesandModulationIndices

As we can see, the simulation results largely agree with the predicted values. The offsets are due to the imperfect filtering. In this section, a mathematical framework was developed to analyze a memristor AM Modulator. It was seen that the results of SPICE simulations showed excellent agreement with the theoretically predicted values. It can be concluded that the AM Modulator behaves according to the mathematical theory both for the non-windowed and the windowed cases. Furthermore, the robustness of the memristor model is shown through the analysis and simulations in this section.

Page48of59

3. Q-Factor Controller for 2ND Order Band-Pass Filter

In this topology, the memristor is used in a 2ND order cascaded Sallen-Key filter topology to set a desired Q-factor. By setting an initial value for the memristor, we can set the Q-factor of the filter to a desired value. By varying the Q-factor, we can change the frequency tolerance of the filter around the center frequency w0. The general transfer function for a 2nd order band-pass filter is:

H ( s) = s2 +

0
Q

Ks s + 0 2

This transfer function can be realized in a cascaded topology using integrator blocks as shown in Figure 33:

1 In1

1 s Integrator

w0 Gain

1 s Integrator1

w0 Gain1

K Gain2

1 Out1

1/Q Gain4

-1 Gain3

Figure 33 2nd order Cascaded Sallen-Key Filter

Page49of59

In this cascaded configuration, the gain of the feedback op-amp is the coefficient 1/Q. Thus, a memristor can be placed here in order to adjust the Q-factor by setting it to an initial value. As long as the input signal is of a high enough frequency, the memristance would remain unaffected, and the filter would be stable at the desired Q-factor. Two different filters are shown in Figure 34. Both are centered on f0 = 10KHz (0 = 62,832Rad/sec). One has a Q-factor of 10, and the other has a Q-factor of 1. In order to achieve unity gain at f0, the filter with Q-factor of 10 needs K=0.1, and the filter with Qfactor of 1 needs K=1.

Q-Factor = 10
0 -5

Magnitude (dB)

-10 -15 -20 -25 -30 -35 -40 3 10

4 5

10

10

Frequency (Hz) Q-Factor=1

0

Magnitude (dB)

-10

-20

-30

-40

-50 2 10

10

10

10

10

Frequency (Hz)

Figure34MATLABAnalysisoftheTransferFunction

Page50of59

The circuit that realizes the filter topology in Figure 33 is shown in Figure 35:

Figure35SPICECircuitforQFactorController

The 0 of this circuit is 1/RC, which is 62,832Rad/sec. An inverting amplifier is placed at the output of the second integrator for two reasons: To provide the negative feedback at the input of the first integrator, and to provide the positive feedback at the input of the second integrator since the memristor gain circuit also an inverting amplifier. The gain K is realized by an inverting gain amplifier where varying RK to changes the gain of the amplifier and inverts back the already inverted output of the second integrator, thus providing the correct polarity at the output. Q-factor of this circuit is one over the gain of the inverting opamp circuit with the memristor. Thus, to achieve the Q-factors of 10 and 1, the gain of the opamp would need to be 0.1 and 1 respectively. To achieve this gain with R2=10K, the memristor would need to be initialized at 1K and 10 K respectively. To find the w0 that will give us the required memristances, we use the memristor equation:

M = RON M = RON

w0 w w w + ROFF 1 0 1K = 100 0 + 16000 1 0 w0 = 0.94 D D D D D w0 w w w + ROFF 1 0 10 K = 100 0 + 16000 1 0 w0 = 0.38 D D D D D

Page51of59

The input signal to both of these filters would need to be at a frequency that is outside the band-pass region. We chose to feed in a sine wave at 100Hz frequency with amplitude of 1V. To get the frequency response of these two circuits, we do an AC sweep on them. The results are shown in Figures 36-37:

Figure36FrequencyResponsewithQFactor=10,K=0.1

Figure37FrequencyResponsewithQFactor=1,K=1

Page52of59

We can conclude that the frequency response of the filter circuit is in full agreement with the mathematically predicted response. We have two degrees of freedom in setting the Q-factor of this circuit: 1- We can vary the w0 between 0 and D, we can achieve any memristance between 100 and 16000. 2- We can choose R2 to be any desired value. Thus, we can conclude that we can design this circuit to have any desired Q-factor.

A valid question to ask is how does this circuit behave in real-time? We can simulate the circuit in real-time and see the result. We would expect w to be largely unaffected since the input signal is outside the band-pass region and is filtered sufficiently before it effects the memristor. Figures 38-40 shows the results of a transient analysis on a filter with Q-factor of 10. Figure X shows w of the memristor as a fraction of D, Figure X shows the voltage at the input of the memristor, and Figure X shows the output of the filter:

Page53of59


Figure38SPICETransientAnalysisofw(t)withQFactor=10

Figure39SPICETransientAnalysisofvoltageatMemristorNodeWithQFactor=10

Page54of59


Figure40SPICETransientAnalysisofOutputVoltageQFactor=10

As we predicted, the w is largely unaffected by the input signal since it is largely filtered out before it reaches the input node of the memristor as can be seen from Figure 39. Furthermore, as can be seen from Figure 40, the input voltage is filtered out at the output by a factor of 0.001 is -60dB, as predicted from Figures 34 and 36 for an input frequency of 100Hz. However, it can also be seen that aside from sinusoidal change, there is a slow but steady downward drift on w. Although this downward drift is negligible in the short-run, in the long-run, this filter would not perform as it was first designed. In fact, as w decreases, the Q-Factor would increase, which would widen the band-pass region. The filter would then let through more of the input signal, which in turn would deteriorate w even more, making the filter more unstable as time goes on. Thus, a correcting

Page55of59

feedback circuit would be necessary in order to keep the w stable. Such a circuit is possible, although it was not analyzed in this thesis. In this section, a mathematical framework was developed to analyze a 2nd order BandPass filter with controllable Q-factor. It was seen that the results of SPICE simulations showed excellent agreement with the theoretically predicted values. It can be concluded that the Q-factor controller circuit behaves according to the mathematical theory. Furthermore, the robustness of the memristor model is shown and reinforced by the analysis and simulations in this section.

Page56of59

VII. Conclusions
Memristors may play a very important role in the future of electronics. It could help us keep up with Moores Law in the digital domain, as applications that use memristors as switches hit the markets within the next five years. However, their most important role will most likely be in the analog domain. By having access to non-volatile analog memory, we can now design new circuits that can do new things. Already, there are patents of applications that use memristors in the fields of signal processing, neural networks, and control systems6. Furthermore, by being able to model neurons without the use of complicated transistor topologies, it may become feasible to built large-scale neuromorphic computers and take advantage of vast parallel-computation, much like a human brain does. To speed up memristor research and achieve these breakthroughs, the problems standing in the way of memristor research must be solved. One of these problems is the lack of software simulation support for these devices. This problem was addressed in the first half of this thesis by creating a behavioral model of TiO2 memristors that can be used to simulate circuits that involve memristors. The results show that the model behaves appropriately and the behavior of the devices reported in [2] are modeled correctly. It is the authors hope that the model described in this paper is used for easily and efficiently simulating circuits that involve TiO2 memristors, and that it serves as a framework for the behavioral modeling of other reported memristive devices, in order to speed up the research and development of memristor electronics.

U.S. Patent 7,302,513; U.S. Patent 7,359,888; U.S. Patent Application 11/976927

Page57of59

The second half of the thesis used this behavioral model to design, simulate, and analyze three circuit topologies: Memristor Switch, Memristor AM Modulator and QFactor Controller for 2nd order Band-Pass Filter. For the Memristor Switch, it was concluded that the memristor switch uses unfeasibly high amounts of energy in order to switch on and off both in the non-windowed case and the windowed case. Further research and explanations by HP Labs should shed light on how memristors can be feasibly used as digital switches. For the AM Modulator, a circuit topology was proposed that uses the memristor in a simple inverting op-amp configuration to modulate a high-frequency carrier. This circuit was analyzed mathematically and SPICE simulation results were compared to theory both for the non-windowed and the windowed memristor. It was concluded that the model behaves according to theory in both cases. Furthermore, performance analysis was done by using looking at THD values. It was concluded that the non-windowed memristor model creates close to zero distortion whereas the windowed memristor introduces distortion as modulation index increases and w(t) approaches the boundaries. In both cases, THD went up as the input signal amplitude decreased, increasing the distortion caused by the high-frequency carrier. For the Q-Factor controller, the memristor was used in the feedback gain section of a cascaded 2nd order Sallen-Key band-pass filter, again in an inverting op-amp configuration. The Q-Factor was controlled by varying the initial condition on the w(t) of the memristor. It was concluded overall that the memristor model created by the author is robust and can be used to simulate new circuit topologies easily and efficiently.
Page58of59

Bibliography
123456Leon, O. C., Memristor - The Missing Circuit Element, IEEE Transactions on Circuit Theory, pp. 507-519, 1971. Strukov D. B., Snider G. S., Stewart D. R., and Williams R. S., The missing memristor found, Nature , pp. 80-83, 2008. Erokhin V., and Fontana M. P., Electrochemically controlled polymeric device: a memristor (and more) found two years ago, Arxiv preprint, 2008. Pershin Y. V., and Massimiliano D. V., Spin memristive systems: Spin memory effects in semiconductor spintronics, Physical Review, 2008. Huai Y., Spin-Transfer Torque MRAM (STT-MRAM): Challenges and Prospects, AAPPS Bulletin, pp. 33, 2008. Borghetti J., Li Z., Straznicky J., Li X., Ohlberg D. A., Wu W., et al., A hybrid nanomemristor/transistor logic circuit capable of self-programming, PNAS, pp. 16991703, 2009. Pershin Y. V., Fontaine S. L., and Ventra M. D., Memristive model of amoeba's learning, Nature Precedings, 2008. Snider G. S., Spike-Timing-Dependent Pearning in Memristive Nanodevices, Nanoscale Architectures , pp. 85-92, 2008.

78-

Page59of59