Abstract
Limit of Detection of Silicon BioFETs
Nitin K. Rajan
Yale University
2013
Over the past decade, silicon nanowire/nanoribbon field-effect transistors (NWFETs) have demonstrated great sensitivity to the detection of biomolecular species, with limits of detection
(LOD) down to femtomolar concentrations. Several well known factors limit the LOD; among them, screening effects, efficiency of the biomolecule-specific surface functionalization, binding kinetics and equilibria, and the delivery of the analyte to the sensor surface. Recently, the noise properties of such biosensors have been receiving more attention, both as a factor that determines the LOD as well as a diagnostic tool to extract information about the electronic properties of the FET sensors.
However, the signal-to-noise ratio (SNR) of these bioFET sensors, and the device parameters that determine the LOD, are not well understood.
We discuss our experiments on applying noise spectroscopy to silicon NWFETs with the goal of understanding and improving the detection limit of such devices. Using low frequency noise measurements and modeling, we are able to compare different devices/material systems and quantify the effect on device performance of different process parameters. We also consider the effects of temperature on the noise generating mechanism and investigate the fundamental origin of 1/f noise in these devices.
We then introduce SNR as a universal performance metric, which includes both the effects of noise and signal transduction, and we show that the SNR is maximized at peak transconductance due to the effects of 1/f noise, and not in the subthreshold regime where sensitivity is maximized.
We also correlate the LOD predicted by the measurement of the SNR to pH sensing experiments, highlighting the relevance of this metric for the ultra-sensitive detection of biomolecules. The effects on the SNR, of surface functionalization, gating scheme and device scaling are also considered and
i quantified, yielding interesting results which will have a profound impact on the design of sensors with lower LOD.
The nanowire-based devices have shown a theoretical LOD of 4 electronic charges, ignoring the effects of screening. Using these devices, with very good performance in terms of SNR, we were able to measure and extract the binding kinetics of protein interactions, which have never been done with NWFETs. Binding constant (KD) determination is a critical parameter for biomolecular design and has until now been primarily assessed by surface plasmon resonance (SPR). The KD determines the magnitude of the sensor signal for a particular concentration of analyte and therefore, is an important factor in determining the smallest measurable concentration. Utilizing the low LOD and reproducibility of sensing signals from our bioFETs, we study the reaction kinetics of low and high
KD systems and demonstrate the viability of the bioFET platform as a potential replacement for
SPR.
Our investigation of different solution gate electrodes and their noise performance show that both the accuracy of biosensing results and the LOD are significantly affected by the choice of the solution gate electrode. A full reference electrode is difficult to integrate into a miniaturized system and pseudo reference electrodes require careful control of the sensing buffer in order to avoid measurement artifacts. We finally propose and demonstrate some intial results on the integration of an on-chip Ag/AgCl pseudo reference electrode for improved noise performance and better LOD, especially under fluid flow conditions.
ii Limit of Detection of Silicon BioFETs
A Dissertation Presented to the Faculty of the Graduate School of Yale University in Candidacy for the Degree of Doctor of Philosophy
by Nitin K. Rajan
Dissertation Director: Mark A. Reed
December 2013 Copyright c 2014 by Nitin K. Rajan
All rights reserved.
iv To my grandfather,
Dharma V. Rajan
v Vitae Non Scholae Discendum
vi Acknowledgments
I acknowledge my advisor, Mark Reed, first and foremost. This thesis would not have been possible without his guidance and constant support, especially when experiments did not work out as planned. I am deeply grateful for the freedom I enjoyed, under Mark’s guidance and encouragement, to explore various ideas. A lot of them did not work out and are not mentioned in this thesis but they resulted in stimulating discussions between Mark and myself, and helped shape my research methodology. This modus operandi of thinking outside the box and trying out different ideas, which is a cornerstone of his research philosophy, kept my inquisitive spark alive and made graduate school both intellectually rewarding and plenty of fun.
I am indebted to Professor T. P. Ma for his suggestions and insightful comments regarding my thesis. He is not only a brilliant scholar but also an excellent teacher, and will serve as a model for my future academic career. Professor Tarek Fahmy has been an invaluable asset and collaborator for the Reed Lab during my Ph.D. and I really hope it stays that way for many years to come.
I am grateful for Tarek’s genuine interest in my thesis as well as his friendship during the past six years. Last but definitely not least, I would like to thank Professor Robert Schoelkopf for his helpful suggestions and unwavering support.
My journey in the academic world would not have been possible without the support of my family. My parents Ashok Rajan and Radha Rajan, made the sacrifice of having their son be two oceans away, so I could study Physics in the US. My father constantly pushed me to develop my critical thinking and taught me, at an early age, to ask questions. My mother taught me
vii to be patient, as she spent countless hours helping me study when I was very young. I am sure the completion of this latest chapter of my life makes them even more proud and they can rest assured that it was all worth it. My sister, Neeha Rajan, deserves a special mention for making me laugh more than anybody else in the world. Graduate school can be hard sometimes, but having somebody with such an iridescent personality in one’s life makes it easier, brighter.
Throughout the years, my path has crossed those of many other graduate students and post- docs. I am extremely thankful for what I learned from them as well as for their friendship. I cannot name them all here, but I have to acknowledge the part they played in my growth as a scholar and more importantly, as a human being. Aleksandar Vacic was a true friend and colleague, who took me under his wing when I first started in the lab and taught me the ropes of the lab which set the foundation for this thesis. He was a great mentor and I always enjoyed the stimulating discussions we had about research, politics, religion and philosophy. As he would like to say, we worked hard and played hard. Weihua Guan’s friendship has been equally invaluable over the years. Within the lab, he is a prolific scientist and I am thankful for the numerous discussions we have had and for what I have been fortunate enough to learn about the process of rigorous scientific thinking. Beyond the lab, the adventures we had made graduate school a very lively period and sometimes even a surreal one. Xuexin Duan and I started collaborating as soon as he joined the lab as a post-doctoral researcher. His scholarship is commendable, his patience is remarkable and his sensitivity makes him a great person to have as a friend. I enjoyed working with him and taking long afternoon breaks discussing science as well as life. Having perspective is extremely important in graduate school and
Xuexin helped me in acquiring some. Another post-doc who was intrumental in my development as a researcher is Eric Jung. His scholarship is well recognized and his simplicity and work ethic are both admirable. I am honored to have known him and I am grateful for his friendship.
Other graduate students at Yale who have been important during my graduate school career, even if our paths crossed only briefly, are: Ryan Munden, David Routenberg, Monika Weber,
Mary Mu, Sylvia Li, Shari Yosinski, Sonya Sawtelle, Zak Kobos, Ben Leung, Xiao Sun, Danti
Chen, Ge Yuan, Daniel Mugaburu, Andrew Zhang, Ricardo Monsalve, Mariana Melo-Vega, Laurie
viii Lomask, Manuel Clemens, Brian Shotwell and Zuhair Khandker. Graduate school has also been an opportunity for mentoring undergraduate students and in the process I have had the honor to work with and eventually built friendships with some truly amazing and brilliant students. They made research more interactive and less lonely, and I am forever thankful to them: Cathy Jan, Jin
Chen, Bo Fan, Shashwat Udit, Vijay Narayan, Paschall Davis and Kara Brower.
None of this thesis would have been possible without the support of many Ithaca College faculty members who I worked closely with and learned so much from. Professors Bruce Thompson, Beth
Clark Joseph, Ali Erkan, Matt Sullivan, Luke Keller, David Brown and Marty Sternstein helped shape my interests and my ability to pursue them, through their teaching, advising and genuine support. My friends, from my college years, deserve a lot of gratitude for being caring and supportive companions during that time: Reuben, Adhish, Elvis, Maksim, Semeret, Innocent, Tendai, Jani,
Goodmore, Doreen, Cathy, Nuha, Alex, Sigurd and Nirbhik.
Finally, I would like to thank the administrative and research support staff at Yale University.
Vivian Smart has brightened so many of my days with her snappy comments and her great sense of humor. She was always genuinely interested in helping me with ordering, shipping or simply bearing with me whenever I needed a break from lab. Arlene Ciocola was always very professional and a pleasure to talk to. The cleanroom staff, namely Michael Power, Chris Tillinghast and Jim
Agresta were always helpful and willing to go the extra step whenever I needed assistance with the cleanroom or with lab equipment and parts. Michael Rooks also deserves a special mention for his assistance and training with the YINQE facilities, as well as his genuine interest in my research projects.
I am very sad to leave the Reed lab, but I am also excited by the current crop of graduate students who without doubt, will be very successful. To you then, whenever the journey gets tough, I hope this quote from Randy Pausch helps:
Experience is what you get when you didn’t get what you wanted. And experience is
often the most valuable thing you have to offer.
ix Contents
List of Figures xiii
List of Tables xxviii
1 Introduction 1
1.1 Overview...... 1
1.2 Performance Metrics...... 4
2 Theoretical Considerations6
2.1 Working Principle of an ISFET...... 6
2.2 pH Sensing...... 9
2.3 Biomolecular Detection...... 12
2.4 Low-Frequency Noise...... 15
3 Device Fabrication and Electrical Characterization 19
3.1 Nanoribbon Device Fabrication...... 19
3.2 Multiplexed Detection Setup Gen.1...... 25
3.3 Multiplexed Detection Setup Gen.2...... 27
3.4 Low Frequency Noise Measurement...... 28
3.5 Lock-in Amplifier Measurements...... 30
3.6 Summary...... 35
x 4 Low Frequency Noise of BioFETs 37
4.1 Introduction...... 37
4.2 Effects of Etching Process on BioFET Characteristics...... 38
4.3 Temperature Dependence of 1/f Noise Mechanisms...... 44
4.4 Random Telegraph Signals...... 49
4.5 Summary...... 58
5 Signal-to-Noise Ratio as a Performance Metric 60
5.1 Introduction...... 60
5.2 Sensitivity...... 61
5.3 Optimal Operating Regime for BioFETs...... 68
5.4 Influence of surface functionalization...... 75
5.5 Gate Coupling...... 78
5.6 Device Scaling...... 81
5.7 Summary...... 87
6 Binding Affinity Considerations 89
6.1 Introduction...... 89
6.2 BioFET as an affinity sensor...... 92
6.3 Binding Kinetics Simulator and SNR...... 93
6.4 High KD Case...... 98
6.5 Low KD Case...... 100
6.6 Summary...... 103
7 Reference Electrode 105
7.1 Introduction...... 105
7.2 Electrode requirements for pH sensing...... 106
7.3 Electrode noise and its effects on SNR...... 112
xi 7.4 Influence of the reference electrode on charge sensing...... 116
7.5 Outlook on integration of on-chip reference electrode...... 118
7.6 Summary...... 121
8 Conclusions 123
References 128
xii List of Figures
1.1 Schematic of a typical bioFET measurement. The baseline current is changed by
the binding of analyte molecules (green cirlces) due to the additional electric field
induced by the charged analyte...... 2
2.1 Schematic of a device and the experimental setup...... 7
2.2 This figure shows an example of how the applied potential Eref is distributed
throughout the different layers of materials involved in the ISFET device structure
as well at the solid-liquid interfaces...... 8
2.3 Circuit model of a typical ISFET sensing setup, showing the voltage drops across
the double layer and silicon device...... 10
3.1 Fabrication Process Flow for Silicon Nanoribbon FET Devices...... 20
3.2 (a) Optical micrograph of a 3.3 mm x 3.3 mm die which is the smallest die we used
in this thesis. (b) Optical image of a larger die size, 6.6 mm x 6.6 mm, which makes
it easier to integrate an external micro-reference electrode into the fluidic cell. The
four on-chip Pt electodes are also clearly visible, extending into the middle of the die. 21
3.3 (a) Optical micrograph of a single nanoribbon device with length and width of 20 µm
and 2 µm respectively. (b) Image of a five-channel device with each channel being
20 µm long and 2 µm wide. (c) A 10-fingered device with L = 10 µm and W = 1 µm. 22
xiii 3.4 Optical micrographs showing devices where the SU-8 passivation windows were not
completely cleared. The problem is more significant for smaller openings. Treating
the devices to oxygen plasma can be used to get rid of the SU8 residues...... 23
3.5 (a) Scanning electron micrograph of a single metallized nanoribbon device, without
the SU-8 passivation layer. (b) SEM image of an array of parallel nanoribbon devices
with the SU-8 passivation layer opening clearly visible in the center. The metal leads
to each device can also be easily seen...... 23
3.6 Scanning electron micrograph of a nanowire device with width of 200 nm and length
of 4 µm...... 24
3.7 Transfer characteristics for a solution gated bioFET device, showing the low threshold
voltage (≈ 1.1V) as well as the low gate leakage current (shown in red and ≈ 20 pA) 24
3.8 Amplifier circuit using an operation amplifier (LT1012) for current to voltage
conversion of the drain current flowing through the bioFET device...... 25
3.9 Photograph of the experimental setup using a connector box, an 8-channel amplifier
stage and a connection panel which interfaces to the NI DAQ card...... 26
3.10 Main Display panel for Labview software designed to concurrently measure 8
recording channels...... 26
3.11 PCB board with assembled components for portable sensing measurements consisting
of on-board amplification channels as well as biasing voltages...... 27
3.12 Gate leakage current measurement circuit for PCB board setup Gen.2. The
sense resistor was chosen such that the circuit is optimized for leakage current
measurements around 1 nA. The voltage after the resistor is measured by Analog
Input (AI) of the DAQ card. The leakage current is then calculated from (Vgs −
Vmeas)/Rsens...... 28
3.13 Modified periodogram method for estimating the power spectrum of a signal. The
algorithm consists of two user defined parameters which are the length of each
periodogram (M) and the degree of overlap (D) between periodograms...... 29
xiv 3.14 (a) 1/f spectra for a bioFET device biased in strong accumulation, interfaced using
the three measurement setups outlined in this section. It is clear that the Portable
setup (Gen.2) performs just as well as the single channel low noise measurement
setup for low frequencies whereas the multi-channel setup (Gen.1) has a very high
noise background. (b) Noise profile for the same bioFET device (extracted at f = 1
Hz) showing the dependence of the noise on the gate bias voltage or drain current.
Again, the portable sensing setup has a similar performance to the single channel
setup except for low current values...... 31
3.15 Normalized noise power spectra of the conductance of a bioFET device as measured
by a lock-in amplifier. The 1/f reference (dashed line) shows that the intrinsic 1/f
noise of the DUT is not affected by the choice of the AC modulation frequency.... 33
3.16 SNR and DC transconductance (gm) as a function of gate voltage in the case of a
DC source-drain bias...... 34
3.17 SNR and AC transconductance (gm, G) as a function of gate voltage in the case of
an AC modulation of the source-drain bias...... 35
3.18 pH calibration curve for a bioFET sensor using the lock-in technique. The AC
current response (proportional to the conductance at Vds = 0) of the PSD is plotted
as function of time, showing the stability of the time traces when an AC source-drain
modulation is used...... 36
4.1 Summary of Hooges parameter (αH ) for different nanowire materials as well as
sub-micron MOS structures utilizing high-k dielectrics. Included is our best SOI
silicon nanowire device (red circle). The dash-dotted line shows the ITRS roadmap
specification of αH for the 45nm technology node (Adapted from [46])...... 39
xv 4.2 SEM images of two nanowires (NWs) etched using an orientation dependent wet etch,
TMAH, and the other one etched using a dry RIE etch chemistry involving Cl2.
TMAH etches (100) planes faster than (111) planes and the resulting trapezoidal
shape can be clearly seen in the SEM image. The device etched using Cl2 has a more
rectangular cross section as expected from an anisotropic RIE etch...... 40
4.3 Plot of noise amplitude, A, as a function of the drain-to-source voltage (Vds) at a
fixed gate voltage, clearly showing that A is independent of the drain current if the
number of charge carriers remains unchanged...... 41
4.4 Typical dependence of 1/f noise spectra on gate voltage for a TMAH-etched device,
from which the noise amplitude A, at each gate voltage (13 - 25 V), can be extracted.
The inset shows 1/A plotted as a function of Vg, where the slope of the line is used
to calculate αH ...... 42
4.5 (a) Measured Hooge parameters for three sets of devices. Each set was etched using
either TMAH or Cl2 or CF4. The box plot shows the 25th percentile, the median, and
the 75th percentile (the mean is indicated by asquare marker). The average values
of αH were 0.0021 for the TMAH devices, 0.015 for the Cl2 devices, and 0.017 for
the CF4 etched devices. (b) Measured subthreshold swing for three sets of devices,
etched using either TMAH or Cl2 or CF4. The average value for the TMAH devices
was 1.0 V/decade. For Cl2 etched devices, the average was 2.6 V/decade, and for
CF4 devices, the average was 3.0 V/decade...... 43
4.6 Typical noise spectra measured for a device at different gate voltages (22 - 40 V)
from subthreshold to strong inversion. The exponents of the 1/ f spectra all lie in
the range 0.8 < β < 1.2. The flattening of the noise spectra is due to background
noise from the measurement setup...... 45
xvi 4.7 Normalized drain current noise amplitude at f=1 Hz (A) is plotted against drain
2 current (Id). Measurements at room temperature (300 K) compared to the gm/I
curve clearly indicate that the device follows the correlated ∆n-∆µ noise model. For
measurements carried out at 100 K, the noise amplitude is consistent with a carrier
number fluctuation noise ( ∆n) model as indicated by the change in slope compared
to the 300 K data in the strong inversion region...... 46
4.8 (a) The data show the gradual change of the slope in the strong inversion regime
as temperature is lowered at T=300, 250, and 100 K, respectively. The data points
for 250 and 100 K have been scaled for easier visualization of the change. (b) The
data are fitted using the correlated ∆n-∆µ noise model, confirming the change in the
noise generating mechanism as the temperature decreases. The Coulomb scattering
coefficient, α, is also extracted from linear fits to the data at 250 and 300 K. The fit
to the data at 100 K clearly indicates that the correlated model is no longer valid at
that temperature...... 47
4.9 Plot of mobility as a function of temperature for 2 NW devices, showing the increase
in mobility as temperature is decreased due to suppressed phonon scattering..... 49
4.10 Evolution of the normalized noise power spectral density from room temperature
(300 K) to 120 K for a NW device. At certain gate voltages and at low temperature
(120 K), only a few traps are active and contribute to the noise spectrum, which
consequently changes from a 1/f spectrum (with the dotted line representing a 1/f
least-squares fit) to a Lorentzian superimposed on a 1/f trend, evident for the larger
frequencies (with the solid line representing a Lorentzian least-squares fit with a
corner frequency of 25 Hz)...... 50
4.11 Illustration of RTS in a current-time trace, showing two discrete current levels with
characteristic times τ0 and τ1, representing the low and high current states respectively. 51
xvii 4.12 Computed results showing how the addition of Lorentzian spectra due to discrete
two-level trapping systems can and indeed does result in 1/f noise spectra for an
ensemble of such traps...... 51
4.13 A segment of drain current versus time measurement for a NW bioFET device
showing the discrete two-level switching, indicating the activity of a single trap
resulting in RTS...... 52
4.14 Normalized noise spectra at different gate voltages showing the typical Lorentzian
spectra associated with RTS. As the voltage is increased to 27 V, it is clearly seen
that the Lorentzian spectrum is changing to a 1/f spectrum, indicating a gate voltage
dependence of the RTS...... 53
4.15 Histogram of current values showing the bimodal distribution characteristic of a two-
level switching signal. The changes in trap occupancy as the gate voltage is varied
can be seen as changes in the populations of the “low” and “high” current states.
From the changes evident in the histograms, we conclude that the active trap in this
case is an acceptor trap, that is the trap is charged when filled and neutral when
empty...... 55
4.16 Plot of the capture (τc) and emission (τe) times as a function of gate voltage at
a temperature of 130 K. The inset shows the typical distribution of times which
follow a Poisson distribution. An exponential or Poisson fit allows us to extract the
characteristic emission and capture times...... 56
4.17 Arrhenius plot of the characteristic emission time (τ0) and capture time (τ1) as
a function of temperature. From the slope, the thermal activation energy of the
capture and emission processes can be extracted.Ea, e = 250 meV and Ea, c = 180
meV...... 56
4.18 Arrhenius plot of the corner freqency (estimated from the noise power spectra) at
different temperatures. The linear fit shows that fc can be modeled by a thermally
activated process with an activation energy of 230 meV...... 57
xviii 4.19 Relative amplitude of RTS plotted as a function of the inverse of drain current
(1/Id). The linear fit shows that the data agrees well with the 1/N dependence of
relative RTS amplitude. The large changes in the level of the RTS noise is attributed
to changes in the scattering coefficient α, resulting in different levels of mobility
scattering...... 58
5.1 (a) Normalized current data for a pH change from pH 7.1 to pH 7.9 at different bias
points (different Id levels). (b) Corresponding Id-Vg curve, where the bias points used
in (a) are indicated with crosses. For measurements done closer to the threshold
voltage, the resulting ∆I/I is larger as can be seen from Equation 5.1 and the
experimental data in (a). The relative noise level is also larger...... 63
5.2 Plot of device reponse (∆I) as a function of the transconductance (gm at different
gate voltage bias) for a pH change from 7.1 to 7.9. From the linear least squares fit,
the pH sensitivity of the device was extracted as 39 mV/pH...... 64
5.3 Illustration of the competitive desorption process of surface bound streptavidin
molecules using a high bulk concentration of D-biotin molecules. The surface bound
streptavidin consists of unbound sites that are available to molecules in the bulk
solution. By introducing D-biotin to the bulk, the strength of that interaction is
enough to pull the streptavidin molecules away from the surface bound biotin molecules. 64
5.4 BioFET response to streptavidin binding (20 pM) in 0.01X PBS, followed by
competitive desorption using 1mM D-biotin. The initial current drop is due to the
positive charge of streptavidin and the subsequent restoration of the baseline current,
on addition of D-biotin, shows that the competitive desorption strategy works. The
sampling rate used was 1000 Hz...... 65
xix 5.5 (a) Normalized current response due to 20 pM of Streptavidin in 0.01X PBS at
different gate voltage values (bias points) for a sampling rate of 1000 samples/s. One
can see that the normalized response is a function of the gate bias, as expected, and
so is the relative noise level (which increases for the lower drain currents). (b) Id-Vg
curve for a biotin functionalized bioFET device, indicating the bias points used for
the sensing measurements in (a) using color coded crosses...... 66
5.6 Plot of the measured signal-to-noise ratio (at a f = 1Hz) as a function of gate bias
voltage. Even though the relative signal change is larger for small overdrive voltages
as can be seen in Figure 5.5, the SNRmeas is lower...... 67
5.7 Plot of the absolute bioFET response as a function of transconductance (gm) at
different gate voltage bias points. From the slope of the linear fit to the data, the
surface potential change of 100 mV can be extracted for 20pM of streptavidin binding
in 0.01X PBS...... 67
5.8 (a) Normalized current noise power density at f= 1Hz is plotted against drain current.
The noise profile does not change significantly with changes in PBS (phosphate
buffered saline) concentration or by changing the electrolyte to KCl (potassium
2 chloride). The proportionality to (gm/Id) confirms that our data are well fitted by
the number fluctuation model. (b) Signal-to-noise ratio (as defined in text) is plotted
against solution gate voltage to highlight the regime at which SNR is maximized. (c)
Transconductance values extracted from I-V measurements are also plotted against
solution gate voltage to point out that maximum SNR occurs close to the point of
peak transconductance...... 69
xx 5.9 (a) Signal-to-noise ratio (as defined in text) is plotted against solution gate voltage
to highlight the regime at which SNR is maximized. The pH of the solution is varied
showing the independence of the peak SNR value on pH, except for the location of
the peak which shifts as the threshold voltage shifts with pH. (b) Transconductance
values extracted from I-V measurements are also plotted against solution gate voltage
to point out that maximum SNR occurs close to the point of peak gm. Peak gm also
shifts with gate voltage due to changes in Vth caused by the different pH...... 71 √ 5.10 The gate voltage noise fluctuations ( SV ) are plotted against solution gate voltage
(limited to the linear regime of operation). The absence of a linear dependence
indicates that number fluctuations are the dominant cause of the noise of these
bioFET devices. The dip in the data also highlights the region of maximum signal- √ to-noise ratio (1/ SV ), which is again shown to occur around the region of peak
transconductance...... 72
5.11 Plot of the signal-to-noise ratio (SNR) and the device transconductance (gm) as a
function of solution gate voltage, highlighting the observation that SNR is maximum
at the point of peak transconductance. The maximum SNR for this device is 11 000
which translates to a minimum detectable pH change of 0.01...... 73
5.12 Current vs time data showing the device response at different pH values. From the
Id-Vg curve and the pH response curve, the sensitivity is determined to be 24.9 mV/pH. 74
5.13 pH sensing experiment to investigate the limit of detection (LOD) of the bioFET
sensor, showing the successful detection of a change in 0.07 pH with a measured
signal-to-noise ratio of ≈ 3.5...... 74
xxi 5.14 (a) Comparison of the normalized current noise power as a function of drain
current, for APTES functionalized devices vs. un-functionalized bioFETs (bare
oxide surface). The APTES functionalization results in a significant reduction in the
current noise power density. (b) The extracted SNR compared for the functionalized
and bare oxide devices, showing the improvement that results from bioFET surfaces
chemically modified with APTES...... 76
5.15 Normalized noise profile for a bare oxide device compared to the same device after
poly-L Lysine (PLL) functionalization. The noise profile is unchanged, which is
consistent with the electrostatic interaction of the positively charged PLL with the
negative silicon oxide surface...... 78
5.16 Optical image of a device showing the PDMS microfluidic channels functionalized
with FITC (green) and channels functionalized with TAMRA (yellow). There is no
leakage of fluorophors between channels as evidenced by the high contrast of the image. 79
5.17 SNR plotted as a function of solution gate bias for different surface functionalization
schemes, demonstrating that the SNR varies minimally for the different types of
surface modifications after the first step of APTES functionalization...... 79
5.18 Measured peak transconductance (gm) plotted for different device dimensions,
showing that gm scales with the ratio W/L as predicted by Equation 5.3. The
R2 value of the linear least squares fit is 0.9947...... 82
5.19 Peak SNR is extracted and plotted for different device dimensions, showing that peak √ SNR scales with WL as expected from Equation. The R2 value of the linear least
squares fit is 0.9988...... 83
5.20 Results from a 2D simulation of the SNR of 2 bioFETs, both with a 50% surface
coverage of bound analyte molecules. The length of the device is increased from 30
to 50 a.u., resulting in an increase in the device area and consequently an increase in
the SNR from 4.7 to 6.1 a.u., which correlates very well with our experimental results. 84
xxii 5.21 Simulation results for the SNR calculated at 50% surface coverage of bound analyte
molecules, for different device areas. The simulation confirms our experimental
findings that SNR is linearly proportional to p(area)...... 85
5.22 SNR as a function of different number of rows and columns, keeping area constant
at 900 a.u. The small differences in SNR show that the layout is not as significant
as the total surface area available for binding...... 86
5.23 Peak SNR extracted for measurements carried out on nanowire bioFETs fabricated
using e-beam lithography, of different widths ranging from 60 nm to 2 µm...... 87
6.1 Langmuir isotherm plotted on a semi-log scale for analyte-receptor systems of
different KD values. The dashed line indicates a certain detection limit which
occurs, for a particular charge per analyte (Z), at a surface coverage of 40%. The
detection limit in terms of concentration is then obviously dependent upon the
binding equilibria of the analyte-receptor system...... 91
6.2 Real-time sensor responses of HMGB1-DNA binding. Each curve represents the
measurement of a different DNA concentration from multiple devices, and sensor
responses are normalized by the transconductance and offset such that all traces
start at zero. The dashed lines represent the linear least squares fit to the data using
Equations 6.2 and 6.3 from which values for kon and koff are estimated...... 94
6.3 Plot of the fitting parameter konρ0 + koff as a function of DNA concentration. The
linear fit indicates good agreement of the data with our binding model and the from
the slope of the line, we can extract the value for kon (i.e. k1)...... 95
6.4 Langmuir isotherm plotted using the data obtained from the real time current traces
of DNA binding to HMGB1. The isotherm is used to extract a value for the KD of
105 nM using Equation 2.24...... 95
xxiii 6.5 Simulation results based on the parameters given in Table 6.1, showing the binding
kinetics for different device areas (and therefore different numbers of receptor sites)
in the case where the analyte concentration (ρ0) is much larger than the KD ..... 97
6.6 Simulation results for binding curves corresponding to different device areas and thus
different numbers of receptors. Here, we consider a high KD system of 1 nM and an
initial bulk analyte concentration of 1 fM, yielding a very low occupation probability
and thus, low numbers of bound molecules at the end-point of detection...... 98
6.7 Coverage ratio extracted as from the binding curves in Figure 6.6 and plotted as a
function of number of receptor sites. Coverage ratio remains constant at 10−6 since
there are always enough molecules in the reaction volume to populate the increasing
number of sites on the surface...... 99
6.8 Coverage ratio extracted from the simulated binding curves for a high KD system
(1 fM) and plotted as a function of number of receptor sites. The coverage ratio
is no longer a constant quantity and instead decreases with increasing the surface
area, since the number of molecules in the bulk are limited and is smaller than that
required for conjugation to the increasing number of receptors...... 101
6.9 SNR as a function of device area, for the case of a low KD (1 fM) and limited
analyte molecules such that the coverage ratio is no longer constant (see Figure 6.8).
This results in a SNR that does not increase with p(area) but instead is
maximized at some intermediate device area. The results are shown for two different
analyte concentration showing that the optimal device size depends on the target
concentration of analyte...... 102
6.10 Coverage ratio as a function of the analyte concentration in the regime of low KD
(1 fM) and limited number of analyte molecules. As the number of receptors is
increased (that is device area is decreased), the 106 receptors curve approaches the
Langmuir isotherm given by Equation 2.24...... 103
xxiv 7.1 pH response of a bioFET device using a Pt wire as the solution gate electrode (pseudo
reference electrode) at two different bias points. The change in the current does not
correspond to changes in the surface charge of the bioFET (TRIS buffer used at
pH 7.45 and pH 7.95), but rather is a measurement artifact due to the interfacial
potential at the reference electrode itself changing with pH...... 107
7.2 Current response (∆I) at different bias points plotted as a function of the
corresponding transconductance (gm), showing that the potential change involved
in the switching of the two well controlled buffers is consistent and repeatable. The
slope of the linear fit line gives a “pH response” of 39 mV/pH...... 108
7.3 Id-Vg curves for two bioFET devices, demonstrating the difference in the interfacial
potential of the Pt electrode as compared to that of the Ag/AgCl electrode. The
larger interfacial potential of the Pt electrodes translates into a more negative
threshold voltage when the transfer characteristics are measured...... 110
7.4 (a) pH response of a bioFET using Ag/AgCl as a pseudo reference electrode and
10mM HEPES as the pH buffer solution. The extracted pH sensitivity is 12.2 mV/pH
(b) pH response of a bioFET using Ag/AgCl and 10mM TRIS as the buffer solution
with an extracted sensitivity of 24.6 mV/pH. This shows the influence of the buffer on
the potential changes that are measured due to the response of the pseudo reference
electrode itself to [Cl–]...... 110
7.5 Id-Vg curves of bioFET sensors for different pH values of the PBS buffer solution,
showing the expected negative Vth shift for smaller pH values...... 111
7.6 Current-time trace for a bioFET device using the back-gate (BG) as a pseudo
reference electrode, showing the lack of response due to the potential change at
the electrode cancelling the change at the device surface...... 112
7.7 SNR as a function of solution gate voltage for a Pt electrode (red) compared to
Ag/AgCl electrode (black)...... 113
xxv 7.8 pH sensing results for the Ag/AgCl electrode under fluid flow conditions of 50
µL/min. The switching spikes correspond to the switching on/off of the electronic
solenoid valves...... 115
7.9 1/f noise spectra for the bioFET biased in strong accumulation, comparing the
spectra of the Pt electrode to the spectra of the Ag/AgCl electrode under no flow
conditions as well as 50 µL/min fluid flow. The spectra for the Pt electrode is noisier
for high frequencies and the noise power increases by a larger amount than the noise
power of the Ag/AgCl electrode, when the flow is switched on as we expected from
earlier measurements...... 115
7.10 1/f spectra for a bioFET biased in strong accumulation, comparing the noise of the
Pt electrode to that of a miniature Ag/AgCl reference electrode, under no flow and
50 µL/min fluid flow conditions...... 116
7.11 Schematic of the experimental setup used to examine the changes in interfacial
potential occuring for various test electrodes upon addition of charged biomolecules,
by meauring the changes in the open circuit potential...... 117
7.12 Open circuit potential (Voc) of the Pt and Ag/AgCl electrodes measured against
a full Ag/AgCl reference electrode, showing the transients and voltage shifts upon
addition of 1 mg/mL of PLL...... 118
7.13 Open circuit potential (Voc) measurement for two full Ag/AgCl reference electrodes
upon addition of 1 mg/mL of Poly-L-Lysine (PLL). The transient is due to the
manual addition of PLL and the voltage change is 0.1 mV...... 119
7.14 The optical micrograph on the left shows patterned Ag metal deposited by e-beam
evaporation on to a silicon substrate. After the treatment with bleach, a thin layer
of AgCl is formed resulting in a darker color. The open circuit potential can be used
to confirm the successful conversion of Ag to AgCl as the potential changes from 200
mV to 36 mV...... 120
xxvi 7.15 Open circuit potential (Voc) measured w.r.t a Ag/AgCl reference electrode as a
function of time, demonstrating the small amount of drift for the on-chip Ag/AgCl
pseudo reference electrode...... 120
7.16 (a) Response of the open circuit potential of the on-chip Ag/AgCl pseudo electrode
(measured w.r.t a reference Ag/AgCl electrode) upon addition of sodium hydroxide
(NaOH), showing that pH changes do not affect the electrode potential as long as [Cl–]
remains constant. (b) Response of the on-chip Ag/AgCl electrode upon an increase
in [Cl–], showing that the electrode potential is indeed responsive to changes in the
chloride concentration...... 121
xxvii List of Tables
5.1 Table showing the average values of the measured SNR and extracted trap densities
for measurements carried out using the top gate silicon oxide, back gate silicon oxide
and top gate aluminum oxide dielectrics...... 81
6.1 Table showing the parameters and their typical values for the binding kinetics
simulations using Matlab’s ODE solver. The area of the bioFET device is given
by A, in units of cm2...... 96
xxviii Chapter 1
Introduction
1.1 Overview
The ability to detect very low concentrations of small molecules such as protein and DNA at a low cost has tremendous applications for medicine and basic biochemistry[1,2,3]. Currently most standard techniques rely on optical characterization methods[4,5] which involve tagging (i.e. labeling) the analyte of interest with a fluorescent molecule. This is not optimal for three main reasons: (1) Attaching a separate molecule to the analyte might influence the latter’s functionality, i.e. the way the analyte binds to other molecules[6]. (2) Pre-processing of samples does not allow the real-time monitoring of changes in analyte concentration and (3) Tagging with fluorophors and reliable optical detection requires a larger sample volume[7]. Alternatively, field-effect sensors such as silicon nanowire/nanoribbon biosensors (BioFETs) have shown great promise as a potential platform for direct, label-free detection of bio-molecules with ultra-high sensitivity and scalability[8,
9, 10, 11]. With the use of micro/nano-fabrication, sensor characteristics can be controlled and true multiplexing and addressing can be achieved. This paves the way for multi-component detection and analysis for Point-of-Care devices utilizing smaller sample volumes and having faster read-outs.
The structure of a BioFET is similar to that of a MOSFET, but instead of having a metal or polysilicon gate electrode, the gate dielectric in the BioFET case is directly exposed to solution.
1 The gate material (usually silicon oxide, silicon nitride or aluminum oxide) is modified with surface receptors specific to the analyte that one wishes to detect. Most biomolecules are charged when in solution, the charge being dependent on the pH. Therefore, when the analytes bind to the receptors, the field-effect due to the charges now bound close to the surface induces opposite charges within the semiconductor channel, which is then detected as a change in current or conductance. The principle of operation of a bioFET is shown in Figure 1.1, where the current rises as the analyte binds to the receptor. The first realization of a bio-sensor utilizing these principles happened in the
1.5µ
ISD (A)
1.0µ SD I
500.0n -50 0 50 100 150 200 Time (sec)
1.5µ
ISD (A)
1.0µ SD I
500.0n -50 0 50 100 150 200 Time (sec)
Figure 1.1: Schematic of a typical bioFET measurement. The baseline current is changed by the binding of analyte molecules (green cirlces) due to the additional electric field induced by the charged analyte.
1970s with the concept of an Ion Sensitive Field Effect Transistor (ISFET)[12]. However, due to the large size of ISFETs, a lot of molecules were needed to achieve a good signal to noise ratio[13] and therefore most of the research on this topic focused on the detection of pH changes. Work carried out on ISFETs eventually led to the pioneering work by Cui et al.[14] which made use of chemical vapor deposition (CVD) to grow silicon nanowires (10 -20nm) and use them as pH sensors as well as bio-sensors to detect proteins (streptavidin) down to picomolar concentrations. Further work on CVD grown semiconductor nanowires showed highly sensitive and real-time detection of
2 proteins[15, 16], DNA[17, 18] and single viruses[19]. However, the bottom-up approach to creating bioFETs suffers from two major drawbacks. Firstly, high density fabrication and integration is complicated by the need to align and/or contact the grown nanowires[20] using unconventional and low yield methods. Secondly, the grown nanowires have a large variance in their physical and electrical characteristics[21]. For true multiplexing capabilities, in the presence of a global gating scheme, the device characteristics need to be similar so that the sensitivity can be optimized for every single device in a multi-device array[13]. The desire to better control device characteristics on a wafer scale led to the top-down approach of using conventional micro-fabrication processes to create nanowire like structures using ultra-thin silicon-on-insulator (UT-SOI) wafers. The devices are defined in the active silicon layer using either optical or e-beam lithography and the rest of the silicon is subsequently etched away using wet-etching[11, 22, 23, 24] or dry RIE etching[10, 25, 26].
Recently, there has also been resurging interest in wider sensors (larger than 1um) with nanometer thicknesses based on SOI, called nanoribbons or nanoplates, which involve a simpler and cheaper fabrication process relying on relaxed-dimension optical lithography. The sensing results from these devices indicate comparable sensitivity to nanowire systems[27, 28], the thickness of the nanoribbons being the critical dimension. Our research group has therefore focused on the top-down approach for better control and uniformity in bioFET characteristics and performance. We have explored both the nanowire as well as the nanoribbon device architectures and have had promising results with both of them. Our study on the detection of the cancer markers PSA and CA15.3 showed the immense potential of having devices with uniform characteristics across a wafer[29]. Due to similar threshold voltages across multiple devices, the devices could be biased at approximately the same operating regime by a single global solution-gate voltage. This allowed for the simultaneous measurement of 8 devices in parallel, greatly increasing the amount of data for a given experiment.
3 1.2 Performance Metrics
The most commonly used performance metric has been the “sensitivity” of the sensor during a certain measurement which is defined as the change in the signal normalized by the original value of the signal (∆I/I) or in other words, the relative signal change. The origin of this metric can be traced back to the pioneering work on nanowire biosensors[14] wherein the authors claimed:
“. . . binding to the surface of a nanowire (NW) or nanotube (NT) can lead to the depletion or accumulation of carriers in the bulk of the nanometer diameter structure (versus only the surface region of a planar device) and increase sensitivity to the point that single-molecule detection is possible.” This argument exploits the higher surface to volume ratio of nanowires as compared to planar ISFETs to justify the increased sensitivity as defined by ∆I/I and led to greater interest in the field of nanowire FET sensors, with single molecular detection being somewhat of a holy grail. Theoretical work attempting to understand and provide guidelines for improving the performance of FET-based sensors therefore focused heavily on modeling how the “sensitivity”
(i.e. the relative signal change) varies with parameters such as bias conditions, doping densities and fluidic considerations[13, 30, 31, 32]. This definition of “sensitivity” should not be confused with the true definition of the sensitivity, which is the change in signal for a certain change in the measurand. In the specific case of bioFETs, measuring changes in the source-drain current,
∆(current) S = (1.1) ∆(analyte concentration)
The change in current can further be expressed as a product of the change in surface potential and the transconductance (gm). We note here that normalizing sensing results by gm has been shown to result in smaller device-to-device variations[29, 33], since the normalized signal becomes solely dependent on the change in surface potential (which is independent of the bias conditions).
Expressing the sensitivity as defined in Equation 1.1 is more relevant to the potential applications of the sensor and allows people to focus on improving the signal transducer which is the bio- recognition layer. The surface coverage density of the functionalization layer as well as the density
4 of active chemical groups in that layer both determine the change in surface potential per unit change in analyte concentration[16]. Defining sensitivity as such will allow more attention to be paid to extracting and displaying the proper calibration curves (the slope of the calibration curve being the sensitivity[34]).
Until very recently[35, 36], little attention had been paid to the fact that the measurement itself is a measurement of ∆I and what ultimately limits the smallest change that can be resolved by the bioFET sensor is the current noise (δi). The latter can be measured, characterized and modeled using established techniques/models in the MOSFET community[37, 38, 39, 40]. Therefore, if the goal is ultra-sensitive detection, the performance metric has to include both the magnitude of the signal as well as the noise properties of the sensor and its environment. In this thesis, we have tried to combine the noise analysis and modeling tools of the semiconductor device community, with the rather novel field of FET biosensing to come up with more relevant and accurate ways to guide the fabrication and/or selection of bioFET sensors with the lowest limit of detection.
With this aim in mind, noise measurement and noise characterization represent valuable tools in being able to predict the smallest measurable signal, or in other words, the detection limit of bioFETs, as well as guiding the fabrication of “better” sensors. In our view, three factors currently determine the limit of detection of a particular sensor: (1) the intrinsic device noise which depends on the fabrication process and material quality, (2) charge screening, especially when physiological conditions (100mM ionic concentration) cannot be avoided and (3) binding equilibria between receptor and analyte molecules which dictate how much of the analyte binds to the surface receptors at steady-state conditions, for a given bulk concentration. This thesis mostly focuses on the first limitation. Chapters 4 and 5 focus on noise characterization of our bioFETs and using
SNR as a performance metric, respectively. Chapter 6 covers binding kinetic measurements carried out using our platform, as well as simulation results based on modeling low and high KD systems.
Finally, Chapter 7 deals with the solution gate electrode and how sensing results can be affected by the choice of the electrode and the buffer composition.
5 Chapter 2
Theoretical Considerations
2.1 Working Principle of an ISFET
To understand the working principle of any FET based biosensor, we need to revisit the work of Bergveld[12]. His invention, the Ion Sensitive Field Effect Transistor (ISFET) is basically a
MOSFET(metal oxide semiconductor field effect transistor) with the gate electrode (solution gate) separated from the gate dielectric by the electrolyte solution. Figure 2.1 shows the device structure for an SOI ISFET, which can be transformed into a bioFET by functionalizing the surface of the dielectric with receptor molecules specific to the analyte of interest. Consequently, any binding event, involving charged analyte species, changes the electrostatics of the system and alters the potential distribution from the solution gate electrode to the ground of the device. The potential distribution for a certain bias voltage Eref is shown in Figure 2.2. Any alteration to this potential distribution is then detected as a change in the drain-to-source current (Ids), which is related to the change in surface potential at the dielectric-solution interface(ψ0).
The drain-to-source current (Ids) in the ohmic region, for low drain-to-source voltage (Vds) is given by[41]: W I = µ C (V − V ) V (2.1) ds L eff ox g t ds
6 Figure 2.1: Schematic of a device and the experimental setup
where W and L are the width and length of the device respectively. µeff is the effective mobility for carriers in the silicon, Cox is the oxide capacitance per unit area, Vg is the applied gate voltage and Vt is the threshold voltage, which is given by:
ΦM ΦSi Qit − Qb + Qf Vt = − − + 2φf (2.2) q q Cox
ΦM and ΦSi are the metal and silicon work functions respectively. Qb, Qf and Qit are the bulk depletion charge, fixed oxide charge and interface charge at the Si/Si02 interface respectively, all per unit area. φf is the fermi potential difference between the doped silicon and instrinsic silicon.
In the case of an ISFET, since the gate voltage is applied via a reference electrode, the threshold voltage (Vt) has to be modified to include additional potential drops at the interface of the reference electrode and the solution and at the interface between the solution and the gate dielectric.
sol ΦSi Qit − Qb + Qf Vt = Eref − ψ0 + χ − − + 2φf (2.3) q Cox
7
Electrolyte Solution Silicon Solution Gate Solution Gate DielectricGate
Eref
φs1 φs2 φox φSi Potential Vb
Distance from gate
Figure 2.2: This figure shows an example of how the applied potential Eref is distributed throughout the different layers of materials involved in the ISFET device structure as well at the solid-liquid interfaces.
This equation is different from Equation 2.2 in that the work function of the metal (ΦM ) is now replaced by the potential applied to the solution by the reference electrode, namely Eref . The potential drop between the reference electrode and the electrolyte is contained within the value of
Eref (In most cases, this potential drop is negligible compared to the applied voltage, and thus,
Vapplied ≈ Eref ) . The additional potential difference at the gate dielectric/solution interface is
sol sol given by ψ0 + χ . χ is the surface dipole moment of the solution which is generally regarded as a constant. ψ0 is the potential contribution due to the charged surface groups on the gate dielectric surface and therefore, ψ0 is a pH dependent quantity and is the essence of how the ISFET can be used as a pH sensor. Before going into the theory behind pH sensing, it is worthwhile to note that bio-molecular sensing can also be included in this model of the ISFET by considering ψ0 to be a function of the number of bound charged analyte molecules. It is also important to keep in mind that even though Equation 2.3 considers the case of a constant Eref , the choice of the reference electrode is crucial in ensuring that Eref is indeed constant during a particular measurement.
8 Otherwise, the ∆ Vt measured cannot be linked solely to changes in ψ0. For instance, in Figure 2.2, a change in current can either be due to the dielectric surface building up charges, resulting in a change in φs2, or the change in current can be due to fouling of the reference electrode producing changes in φs1.
2.2 pH Sensing
The theory of pH sensing is based on the site-binding/site-dissociation model[42] and involves the following surface hydroxyl groups and their corresponding reactions:
Ka − + A−OH )−−−−* A−O + Hs (2.4)
+ Kb + A−OH + Hs )−−−−* A−OH2 (2.5)
+ The concentration of hydrogen ions directly near the surface ([H ]s) determines the state of the hydroxyl groups on that surface. The surface concentration of hydrogen ions is related to the bulk concentration by: q ψ [H+] = [H+] exp(− dl ) (2.6) s b k T
+ where [H ]b is the bulk concentration of hydrogen ions, k is the Boltzmann constant, T is the temperature and ψdl is the potential difference between the dielectric surface and the bulk of the solution. A circuit model of the sensing experiment is shown in Figure 2.3. The bulk of the silicon is assumed to be grounded at some point (usually at the source contact) and ψ0 is the potential at the gate dielectric surface. This potential is subject to change when the charge at the dielectric surface (Q0) changes. The charge separation in the double layer at the interface of the oxide and electrolyte is modeled as a simple capacitor, Cdl. Figure 2.3 illustrates the charge sharing picture of charges binding at the surface of the gate dielectric. The surface charge is compensated by charges in the double layer as well as charges in the silicon. Under such an approximation, the potential
9 Figure 2.3: Circuit model of a typical ISFET sensing setup, showing the voltage drops across the double layer and silicon device.
difference, ψdl can be related to the oxide surface charge per unit area, σ0, by the following:
Cox Cnw σ0 = σdl + σnw = Cdl ψdl + ψ0 (2.7) Cox + Cnw
where Cnw is the capacitance of the nanowire, which is only significant in the presence of a depletion region within the silicon. Henceforth, in order to simplify the mathematics, we will consider the case where Cnw is much larger than Cox, which is often the case for our measurements which are carried out in the strong accumulation or strong inversion regimes. Rewriting ψ0 as a function of the applied solution gate voltage, Vsg, we obtain:
σ0 = ψdl(Cdl + Cox) + Vsg Cox (2.8)
We can alternatively write σ0 as a function of ψ0 which yields the following:
σ0 = ψ0(Cdl + Cox) − Vsg Cdl (2.9)
σ0 is also related to the areal density of charged surface groups:
+ − σ0 = q([A−OH2 ]−[A−O ]) (2.10)
10 where the total areal density of surface groups is given by:
+ − Ns = [A−OH2 ] + [A−OH] + [A−O ] (2.11)
Taking the logarithm on both sides of Equation 2.6 and combining with Equations 2.10 and 2.11, we obtain the following equation:
1 1 2 2 + Ka qψdl −1 σ0 1 ln [H ]b − ln = + sinh (2.12) Kb kT qNs 4KaKb
+ Re-writing Equation 2.12 using the relation pH = − log10[H ]b,
1 2 qψdl −1 σ0 1 2.303 (pHpzc−pH) = + sinh (2.13) kT qNs 4KaKb
where pHpzc is the pH value at the point of zero charge(pzc), that is when the surface of the oxide is electrically neutral (σ0 = 0 and ψdl = 0). Equation 2.13 can be combined with Equation 2.8 to give the following general equation in the presence of charge sharing and applied solution gate bias:
qψdl −1 q Cox 2.303 (pHpzc−pH) = + sinh 0 ψdl + Vsg (2.14) kT kT β Cox + Cdl
0 where β is given by: 2 1 0 2q N (K K ) 2 β = s a b (2.15) kT (Cdl + Cox)
0 It is evident from Equation 2.15 that β , being proportional to the areal density of surface groups and how easily these groups associate or dissociate to give a charged surface, relates to the reactivity
0 of the surface. For a good pH sensing surface, we expect β to be large, more specifically, we expect
0 qψdl β kT , which results in the following approximation,
0 kT β Cox ψdl = 2.303 0 pHpzc−pH − Vsg 0 (2.16) q β + 1 (Cox + Cdl)(β + 1)
11 The surface potential ψ0 which is given by Vsg + ψdl can now be defined, resulting in the Nernst equation for pH sensing, with the solution gate bias taken into account:
0 kT β Cox ψ0 = 2.303 0 pHpzc−pH + Vsg 1 − 0 (2.17) q β + 1 (Cox + Cdl)(β + 1)
0 For a large value of β , a unit change in the pH value produces a surface potential change of 59.9 mV, which is called the Nernst limit. This value is not possible with silicon dioxide as the sensing surface but is routinely achieved with other gate dielectrics such as aluminum oxide and tantalum oxide. It is important to note that this is a fundamental limit of a sensing surface and despite many attempts to go beyong this limit, the only successful strategy involves amplification of the signal giving an ”apparent” surface sensitivity that exceeds the Nernst limit. The fundamental limit to the surface sensitivity, however, remains unchanged.
2.3 Biomolecular Detection
The detection of charged molecules can be described using an analogous model to the one developed in the previous section. The main difference to note is that with charge detection we consider the charges induced in the silicon channel as opposed to the potential set-up at the surface of the gate dielectric. In principle, both approaches should yield the same results except for the effects of screening. In the case of pH sensing, the reactive groups (hydroxyl groups) are close to the surface and within the Debye length for most electrolyte concentrations used. However, when trying to detect charged analyte molecules, the surface has to be modified and functionalized with specific receptors, which have lengths of about 10 nm or longer. In such a case, the Debye length has to be optimized by choosing the right ionic concentration for the buffer solution.
s r0kT λD = 2 (2.18) 2NAq I
12 where r is the relative permittivity of the electrolyte, 0 is the vacuum permittivity, NA is
Avogadro’s number and I is the ionic concentration in mole per m3. For a certain charge Q at a certain distance d (due to the thickness of the functionalization layer), the effective charge
Qeff that can be detected by the bio-sensor is given by:
−d Q0 = Q exp (2.19) λD
This charge Q0 is then shared as before between the double layer on the solution side and the semiconductor channel on the device side, as depicted in Figure 2.3. From Equation 2.9, we conclude that any binding event which changes the surface charge, in turn changes the surface potential ψ0 and these two quantities are related to each other by:
1 ∆ψ0 = ∆σ0 (2.20) Cdl + Cox
We have described how the surface potential of the device changes after a certain number of receptors are conjugated. The physics of this conjugation is quite complex if one includes the effects of diffusion, mass-transport, non-specific binding, multiple binding sites, etc. For the sake of simplicity, a lot of insight can be obtained by considering the case of one-to-one binding (which is generally what is sought after in the design of bioFETs) with the addition of a two compartment model[43], to distinguish between reaction-limited and diffusion-limited regimes.The basic differential equation governing the binding kinetics of a receptor-analyte pair is given below:
∂N = k ρ (N − N) − k N (2.21) ∂t on s max off
where N is the number of conjugated receptors, Nmax is the total number of receptors available, kon and koff are the association and dissociation rate constants, respectively. ρs is the local concentration (in mol/L) of analyte molecules, right above the layer of receptors, to be distinguished from the bulk concentration of analyte, ρ0, far away from the reaction layer(reaction zone). The
13 two-compartment model was developed to analyze the depletion of analyte molecules in the reaction zone and the subsequent diffusion of analytes from the bulk into that region or compartment. This diffusion is modeled by a diffusion rate constant, km and is represented by the following differential equation: ∂ρs 1 = [km (ρ0 − ρs) − konρs (Nmax − N) + koff N] (2.22) ∂t VNA
where V represents the volume of the reaction zone in L and NA is the Avogadro number. We can consider the case of a reaction-limited experiment, which means that the reaction proceeds slowly enough such that the local concentration is always equal to the bulk concentration (ρs = ρ0). In such
dN a case, at equilibrium ( dt = 0), Equation 2.21 produces a simple equation for N at equilibrium:
kon ρ0 Nmax Neq = (2.23) koff + kon ρ0
koff The above equation can be re-written in terms of the equilibrium dissociation constant KD = , kon
ρ0 Neq = Nmax (2.24) ρ0 + KD
Therefore, for a given bulk analyte concentration ρ0 and knowing the equilibrium dissociation constant KD, the equilibrium number of bound molecules can be estimated by using Equation 2.24, assuming the conjugation is reaction-limited. Otherwise, Equations 2.21 and 2.22 need to be solved numerically to figure out Neq. After Neq is determined, the bound surface charge(Qbound) can be estimated by knowing the charge per molecule, which is determined from the iso-electronic point
(pI) of the analyte molecule and the pH of the electrolyte. Screening can then be accounted for using Equation 2.19 and the change in surface potential (ψ0) can be formulated as:
1 Z ρ0 Ns −d ∆ψ0 = exp (2.25) Cdl + Cox KD + ρ0 λD
14 where Z is the charge per analyte molecule, which is pH dependent and Ns is the areal density of surface receptor molecules, which is related to the efficiency of the functionalization. Ns = Nmax/A where A is the area of the gate dielectric surface functionalized with receptors and exposed to the solution containing analytes. It is very important to note that Equation 2.25 was derived by assuming that ρs is constant and equal to ρ0, which is true in most analyte-receptor systems that are reaction-limited, that is with high KD (>1nM). In the case of strong binding systems (low
KD, <1pM), the number of surface bound molecules at equilibrium does not necessarily follow
Equation 2.24 since ρs is time-dependent and decreases as more analyte molecules bind to the surface.
2.4 Low-Frequency Noise
The smallest measurable signal is determined by the noise level of the bioFET sensor. Since the device is essentially a MOS (metal-oxide-semiconductor) transistor, the noise at low frequencies (<
10kHz) is dominated by flicker or 1/f noise[39][44] which has the following general form:
S A I = (2.26) I2 f
2 SI is the power spectral density of the drain current noise (in A /Hz), I is the drain current and A is a constant of proportionality which differs based on the noise model that is used. The quantity
2 SI /I is the normalized noise power density (the noise power is normalized by the power dissipated
2 by the device, assuming a resistance of 1 ohm) and at a frequency of 1Hz, A = SI /I and is defined as the noise amplitude. The noise amplitude is the only quantity necessary to define the
1/f noise spectrum completely and is the quantity that contains the details of the noise generating mechanisms and properties. Hooge’s model for instance[45], is based on an empirical equation, where A = αH /N, such that: S α I = H (2.27) I2 f N
15 where N is the number of charge carriers in the material and αH is called Hooge’s parameter and is material specific. αH is dimensionless and is a measure of the quality or equivalently the
“noisiness” of the material/device. Typical values are in the range 10−6 to 10−3. This model is commonly used in the linear (ohmic) region of a FET device and helps to compare across different device architectures and materials[46]. To represent devices and the drain current noise from the subthreshold regime to the strong accumulation/inversion regime, the number fluctuation (∆N) model is more accurate. The ∆N model is a physical model which assumes that the origin of the current fluctuations is the trapping/de-trapping process of charge carriers tunneling to/from traps and defect levels in the oxide close to the semiconductor interface. These fluctuations in the number of charge carriers in the channel are modeled as fluctuations in the flatband voltage:
2 λ k T q Not SVFB = 2 (2.28) WLCox f where λ is the characteristic tunneling distance, which is material dependent (for silicon oxide, it is on the order ot 10−10 m). This model assumes a spatially uniform trap density within a few kT
−1 −3 of the Fermi level. Not is the oxide trap density in units of eV m . Therefore, SVFB is usually constant for a given device at different drain currents, unless the trap distribution, Not, varies with gate voltage. The normalized drain current fluctuations are then given by:
S g2 I = m S (2.29) I2 I2 VFB
2 Equation 2.29 accurately models the noise profile (plot of SI /I vs. I) in the subthreshold regime, where gm/I is a constant, resulting in a flat region (since SVFB is usually weakly dependent on the gate bias voltage). The ∆N model predicts a normalized noise power density that varies as
2 1/I in the strong inversion/accumulation regime (for a constant gm). However, many devices do not display this trend. In fact, the dependence varies between 1/I2 and 1/I[47][48], that is the actual data lies somewhere between what is predicted by the ∆N and ∆µ models. To account for
16 this, a third model was developed, called the correlated number-mobility fluctuation (∆N − ∆µ) model[47]. The physical basis for this model is that trapped charges not only influence the number of charges in the channel, but also influence the mobility, since charged traps also act as Coulombic scattering centers which result in mobility fluctuations as the trap occupancy fluctuates. In the
∆N − ∆µ model, the gate voltage noise power density is given by:
p p SV g = SVFB [1 ± α µeff Cox (Vg − Vt)] (2.30)
where α is the Coulombic scattering coefficient and the plus or minus signs depend on whether the trap is charged or neutral, respectively, when it is occupied. For α = 0, we recover the number
fluctuation model of Equation 2.29, with SV g = SVFB. For a non-zero α, the deviation from the number fluctuation model becomes more pronounced in the strong inversion/accumulation region, that is when the gate overdrive voltage (Vg − Vt) is large. In such a case, the normalized noise
2 β power increases for large drain current values and SI /I ∼ 1/I where 1 < β < 2.
Any bio-sensing experiment, for a given analyte-receptor system at a certain concentration, results in a change in the surface potential (ψ0) given by Equation 2.25. The noise (δi) in such a measurement, if the noise spectrum follows a 1/f spectrum, is calculated as:
Z f2 Z f2 2 SI (f = 1Hz) f2 (δi) = SI df = df = ln × SI (f = 1Hz) (2.31) f1 f1 f f1
Consequently, the signal-to-noise ratio of such a measurement (SNRmeas) can be defined as:
∆I gm × ∆ψ0 SNRmeas = = √ p (2.32) δi BW SI (f = 1Hz)
where SI (f=1Hz) is the drain current noise power spectral density at f=1Hz and BW is a bandwidth related term which depends on the largest (f2) and smallest (f1) frequencies sampled, BW = √ ln f2 . For typical sensing measurements, BW is a weak function of the measurement bandwidth f1 and ranges from about 3.5 to 3.8 It is obvious that the smallest ∆ψ0 that can be measured will
17 have SNRmeas = 1. Of course, for reliable detection, we usually consider SNRmeas ≥ 3, but since we are discussing the limit of detection, it makes sense to consider the smallest possible SNRmeas which results in an observable signal. With SNRmeas = 1, we are led to define a metric that we shall name, for lack of a better term, SNR,
gm 1 SNR = p = √ (2.33) SI (f = 1Hz) SVFB
√ Equation 2.33 defines SNR as the SNRmeas for a unit voltage change in ψ0 and for BW = 1. Based on this definition of SNR, it is clear that the limit of detection (LOD) is given by 1/SNR, which
p √ is the gate voltage referred noise amplitude ( SV g or SVFB). The LOD can thus be extracted via a combination of DC and noise measurements, even before any actual sensing experiments are carried out. In this thesis, we will make the case for using SNR as a performance metric which takes into account the signal transduction as well as the noise limitations of bioFET sensors.
18 Chapter 3
Device Fabrication and Electrical
Characterization
3.1 Nanoribbon Device Fabrication
Improving on earlier generations of nanowire/nanoribbon FETs[49][50], this fabrication process includes a top-gate oxidation step (alternatively this could be replaced by an atomic layer deposition step) which greatly improves device lifetime and stability. The passivation layer was also changed back to SU-8, to allow the use of organic solvents during the biofunctionalization procedure. A passivation layer of around 2µm was found to be stable against electrolyte diffusion and protected the devices and the metal leads for several weeks. The general process flow is shown in Figure 3.1.
We start with a 4-inch silicon-on-insulator (SOI) wafer from SOITEC (Active layer: 70 nm, P-Type
10 Ω cm, Buried Oxide Layer: 145 nm, Handle Layer: 500 µm, P-Type 15 Ω cm). Following RCA cleaning of the wafer (SC-1 and SC-2), the first step is to thin down the active layer using sacrificial oxidation, followed by silicon oxide removal using dilute hydrofluoric acid (HF:H2O, 1:10). The
final active layer thickness is chosen such that it is close to the Debye length of silicon at that
15 doping concentration (for Na = 1x10 , λDebye = 40 nm) in order to maximize sensitivity[28][51].
19 Active Si BOX 1. Thin wafer Si Handle Implant TOX 2. Etch alignment marks Al SU-8
3. Ion implantation
4. Mesa definition
5. Oxidation
6. Metallize and liftoff
7. Passivate w/ SU-8
Figure 3.1: Fabrication Process Flow for Silicon Nanoribbon FET Devices
Step 2 involves patterning and etching the alignment marks into the handle silicon layer, using
SF6O2/C4F8 chemistry which etches silicon at a very fast rate (∼ 600 nm/min) and also etches silicon oxide (∼ 120 nm/min). In order to obtain a device with ohmic contacts, the source and
+ drain regions are implanted in Step 3 at the Cornell Nanofabrication Facility using BF3 at 10 keV for a final dose of 2 − 5 x 1015 cm−2. Not shown in Figure 3.1 is the etching of the back-gate contacts through the active silicon and the BOX, which are also implanted. The dopant atoms
◦ are then thermally activated by heating the wafer at 900 C for 10 minutes in a N2 atmosphere.
Following dopant activation, the devices are defined in the active silicon layer (Step 4) using optical lithography (minimum device width = 1µm) and subsequent etching using an Inductively Coupled
◦ Plasma (ICP) etch with Cl2. The devices are then oxidized at 1100 C in a dry O2 environment to create a top oxide (TOX) of about 20 nm as shown in Step 5. Aluminum contacts and leads are then patterned via optical lithography followed by a brief buffered oxide etch (BOE) to get rid of the top-oxide in the source/drain contact regions. The aluminum is then evaporated using an
20 (a) (b)
Figure 3.2: (a) Optical micrograph of a 3.3 mm x 3.3 mm die which is the smallest die we used in this thesis. (b) Optical image of a larger die size, 6.6 mm x 6.6 mm, which makes it easier to integrate an external micro-reference electrode into the fluidic cell. The four on-chip Pt electodes are also clearly visible, extending into the middle of the die. electron beam evaporation system and the final pattern is obtained by a lift-off process resulting in an aluminum thickness of 200 nm (Step 6). Not shown in the process flow of Figure 3.1 is a second metallization step for the on-chip pseudo-reference electrode (consisting of 5 nm of Ti followed by 50 nm of Pt). As will be discussed in Chapter 7, the on-chip Pt electrode was found to lead to erroneous results and was therefore initially abandoned for an external reference electrode and has now been modified to be a silver metal (Ag) layer which is to be further oxidized to AgCl. To improve the quality of the metal-semiconductor contact, thermal annealing is carried out in a tube furnace at
◦ 470 C, in a N2 atmosphere for 10-15 minutes for the formation of ohmic contacts and in order to reduce the contact resistance. Longer times and slightly higher temperatures might be required if the current-voltage characteristics are not linear and the contact resistance is much higher than about 1 kΩ, as measured between back-gate contacts. The final step (Step 7) is the passivation of the metal leads such that they are not exposed to solution. This is achieved by patterning a 2
µm-thick SU-8 layer (hard baked at a final temperature of 150 ◦C) with openings on top of the device channels. Optical micrographs of completed dies are shown in Figure 3.2. Each fabricated wafer consists of single silicon channels between the source-drain contacts as well as multiple silicon channels with common source and drain contacts (see Figure 3.3). These “multi-fingered” devices
21 (a) (b)
(c)
Figure 3.3: (a) Optical micrograph of a single nanoribbon device with length and width of 20 µm and 2 µm respectively. (b) Image of a five-channel device with each channel being 20 µm long and 2 µm wide. (c) A 10-fingered device with L = 10 µm and W = 1 µm. were used for the studies on device area dependence and were found to be more robust that the single devices, which we believe to be due to the larger separation between the metal leads and the edge of the passivation window opening. Figure 3.3 also shows the passivation window opening in the SU-8 layer more clearly. The SU-8 processing can be a little tricky and it is very important to ensure that firstly, the SU-8 is not over-exposed and secondly, that the wafer is developed with significant agitation to ensure that the smaller openings are completely cleared. Figure 3.4 shows devices where the SU-8 windows were not completely cleared. In order to get rid of the residual
SU-8, the devices had to be exposed to oxygen plasma for roughly 4-5 minutes at 220W (or until the windows were visibly clean of any SU8 residues). Over-exposure to oxygen plasma results in a very thin SU-8 passivation layer which leads to significant leakage current from the metal leads to the solution gate electrode. Therefore, care should be taken when carrying out this step.
Scanning electron microscope images of completed nanoribbon devices are shown in Figure 3.5, showing very smooth sidewalls which are very important for low-noise and high transconductance devices. Similar sidewall characteristics were obtained for nanowire devices as shown in Figure 3.6.
These devices were fabricated using electron beam (e-beam) lithography at the Cornell Nanofabri- cation Facility (CNF). The measurements carried out in this thesis involve both measurements on nanowire as well as nanoribbon devices. The nanowire devices were used for the dry, temperature based measurements described in Chapter 4 and the nanoribbon devices used for solution gated
22 (a) (b)
Figure 3.4: Optical micrographs showing devices where the SU-8 passivation windows were not completely cleared. The problem is more significant for smaller openings. Treating the devices to oxygen plasma can be used to get rid of the SU8 residues.
(a) (b)
Figure 3.5: (a) Scanning electron micrograph of a single metallized nanoribbon device, without the SU-8 passivation layer. (b) SEM image of an array of parallel nanoribbon devices with the SU-8 passivation layer opening clearly visible in the center. The metal leads to each device can also be easily seen. measurements of Chapter 5. Based on previous sensing results[52], we obtain very good sensitivity with the nanoribbon devices and additionally the device characteristics are more uniform for the optically defined nanoribbon FETs[29]. We also found that the optical lithography process (i.e nanoribbon fabrication) resulted in more robust bioFET devices, with low turn-on voltages (between
0V and -5V) and low leakage currents (∼ 30pA) when exposed to high salt concentrations (100 mM
NaCl). Figure 3.7 shows a typical Id-Vg curve for a solution gated nanoribbon device showing the low leakage current as well as the low turn-on voltage.
23 Figure 3.6: Scanning electron micrograph of a nanowire device with width of 200 nm and length of 4 µm.
Figure 3.7: Transfer characteristics for a solution gated bioFET device, showing the low threshold voltage (≈ 1.1V) as well as the low gate leakage current (shown in red and ≈ 20 pA)
24 RGain= 1Mohm
Vds DUT - + Vout
Vg
Figure 3.8: Amplifier circuit using an operation amplifier (LT1012) for current to voltage conversion of the drain current flowing through the bioFET device.
3.2 Multiplexed Detection Setup Gen.1
After the wafer fabrication, the wafer is diced into individual dies, with each die consisting of a number of devices (from 14 devices upto a maximum of 32) that are individually addressable.
Therefore, in order to get the most out of our sensing measurements and verify the repeatability of these measurements, we need to have a way to monitor a number of devices concurrently. Thus, we developed a multiplexed detection setup to concurrently measure 8 devices. Each amplification channel consists of an operational amplifier configured as a current-to-voltage converter which feeds into a National Instruments Data Acquisition (NI DAQ) Card as depicted in Figure 3.8. The gain resistor (Rgain = 1 MΩ) was chosen to accomodate typical drain currents of ∼100 nA such that the gain would be 106 (1 µA of input current produces 1 V of output voltage). In order to interface with the electrical measurement setup, the die has to be wirebonded and packaged in a ceramic
DIP holder (Spectrum Semiconductor Inc.). The device can then be hooked up to a connector box, amplifier box and DAQ card as shown in Figure 3.9. The measurement of 8 channels is then carried out and recorded using a Labview program (designed in-house) as can be seen in Figure 3.10.
Concurrent sensing measurement results, using this multiplexed detection setup, have been shown previously[29].
25 Connector Box Op-Amp I-V Converter Box DAQ Card
Figure 3.9: Photograph of the experimental setup using a connector box, an 8-channel amplifier stage and a connection panel which interfaces to the NI DAQ card.
Figure 3.10: Main Display panel for Labview software designed to concurrently measure 8 recording channels.
26 Figure 3.11: PCB board with assembled components for portable sensing measurements consisting of on-board amplification channels as well as biasing voltages.
3.3 Multiplexed Detection Setup Gen.2
In order to further improve the portability and the noise performance of the measurement setup, the amplification stage was combined with the connector box (interfacing to the DIP holder) in a custom designed printed circuit board (PCB). The PCB design involved eight amplification channels, based on the same circuit as in Figure 3.8, with on-board switches for individual channels and individual op-amps. The board was connected to the DAQ card via a 68-pin VHDCI connector from NI and custom Labview software was written to interface and extract data from the 8 channels concurrently.
The final assembled PCB board is shown in Figure 3.11. The board supplies both the drain-to- source voltage (Vds) as well as the gate-to-source voltage (Vgs). There is an external BNC connector and an associated switch to allow for an external gate voltage source since the on-board voltages are limited to 10V. There is also a gate leakage current measurement circuit which relies on a shunt resistor to convert the current into a voltage that can be subsequently measured by the analog-to-digital converter (ADC) of the NI DAQ card. The schematic of the leakage current
27 - AI 31 +
AO 1
(VGS) RSens= 100Mohm
Figure 3.12: Gate leakage current measurement circuit for PCB board setup Gen.2. The sense resistor was chosen such that the circuit is optimized for leakage current measurements around 1 nA. The voltage after the resistor is measured by Analog Input (AI) of the DAQ card. The leakage current is then calculated from (Vgs − Vmeas)/Rsens. measurement circuit is shown in Figure 3.12. The smallest leakage current than can be measured by this circuit is around 100 pA. The voltage after the sense resistor (Rsens ) is measured by one of the analog inputs (AI) of the DAQ card and the voltage drop is used to calculate the leakage current, Ileak = (Vgs − Vmeas)/Rsens, which is then used to determine whether a certain bioFET can be used.
3.4 Low Frequency Noise Measurement
We utilize the same data acquisition capabilities outlined in the previous sections to measure a large number of drain current values at high frequencies (typically 1000 Hz). To focus on the noise which makes up the fluctuations in the drain current, the measured signal is first filtered and then amplified using a low noise current amplifier from Stanford Research Systems (SR570). The bandpass filter used has corner frequencies of 0.03 Hz and 1000 Hz, which gets rid of the equilibrium DC signal and allows us to measure and record only the AC fluctuations in the current, that is the noise signal, within the bandwidth set by the bandpass filter. The output of the current amplifier is a voltage which can be fed into the NI DAQ card, interfaced again via a custom designed Labview program.
Noise analysis involves converting a time-domain signal into a signal in the frequency domain which essentially is a Fourier transform. The algorithm we use for our noise power spectrum extraction is
28 Figure 3.13: Modified periodogram method for estimating the power spectrum of a signal. The algorithm consists of two user defined parameters which are the length of each periodogram (M) and the degree of overlap (D) between periodograms
Matlab’s pwelch method[53] which is a modified periodogram method for estimating the frequency domain power spectrum. Basically, the full signal is split up into smaller segments of data, based on what the user selects as the size of the periodogram (M), and an FFT is then carried out on each segment and finally averaged to obtain the power spectrum of the whole data set. The basic principles of the algorithm are depicted in Figure 3.13. The periodogram size (M) and the degree of overlap (D) between them, determines the number of periodograms generated. More periodograms implies smaller variance in the final power spectrum that the algorithm outputs whereas smaller sized periodograms implies that the frequency resolution of the final power spectrum is low.
Using our noise characterization tools, we can compare the noise performance of the different measurement setups, while measuring the noise profile of low noise device. The results of such a comparison are showin in Figure 3.14. It is evident that the portable sensing setup (Gen 2.) has a
29 similar performance to our low noise single channel measurement setup, which is as expected since the amplification stage is closer to the devices. It is also important to note that the portable sensing setup fares worse when the frequency is high, due to the larger background noise of the op-amps we utilized in our PCB design. Also, for low current conditions, the noise amplitude is larger in the case of the Gen. 2 setup which is again most likely due to the higher noise contributions of the op-amps in our circuit. The multi-channel setup (i.e. Gen. 1) has a higher background noise level is general, though the effect is more pronounced for low drain current.
3.5 Lock-in Amplifier Measurements
The lock-in amplifier is an instrument commonly used to make very sensitive, low-noise measure- ments. The working principle is that instead of measuring the DC response of a device where the
1/f noise contribution from other electronic components in the circuit is significant, one measures the AC response of the device at a frequency which minimizes the noise contributions from those components. The voltage applied to the device under test (DUT) consists of a small AC modulation superimposed on a DC bias voltage:
Vin = VDC + vac sin(ωt + φsig) (3.1)
This results in a device current which also consists of a small AC current superimposed on a DC current level: d I I = I + v sin(ωt + φ ) (3.2) out DC d V ac sig where the term d I/d V is the conductance of the device. 1/f noise fluctuations of the device itself can be modeled as fluctuations in the conductance, G:
d I = G + δg(t) (3.3) d V
30 -2 10 (a) -4 10
-6 10 (1/Hz) 2 /I
I -8
S 10 Low Noise -10 10 Multi-Channel Portable-Sens
-12 10 -1 0 1 2 10 10 10 10 Frequency (Hz) 0 10 (b) Low-Noise Multi-Channel -2 Portable-Sens 10
-4 10 Noise Amplitude Noise -6 10
-8 10 -10 -9 -8 -7 -6 10 10 10 10 10 Drain Current (A)
Figure 3.14: (a) 1/f spectra for a bioFET device biased in strong accumulation, interfaced using the three measurement setups outlined in this section. It is clear that the Portable setup (Gen.2) performs just as well as the single channel low noise measurement setup for low frequencies whereas the multi-channel setup (Gen.1) has a very high noise background. (b) Noise profile for the same bioFET device (extracted at f = 1 Hz) showing the dependence of the noise on the gate bias voltage or drain current. Again, the portable sensing setup has a similar performance to the single channel setup except for low current values.
31 The phase sensitive detector (PSD) of the lock-in amplifier then essentially multiplies the current signal from the device (Iout) by a reference sine signal of the same frequency sin(ωt + φref ) and a low pass filter is subsequently applied (which gets rid of the non DC signals). The PSD output is then given by:
PSDout = vac cos(θ)[G + δg(t)] (3.4)
where θ = φsig − φref . Usually, the phase of the reference signal is adjusted to make the cosine term equal to 1. The output of the lock-in amplifier is then the AC current response of the device, vac × G. However, it is evident from Equation 3.4 that the output of the PSD includes the conductance fluctuations of the device, δg(t), which can be attenuated by the low pass filtering and this attenuation is no different than that obtained by low pass filtering in the case of a traditional
DC measurement. The normalized noise power spectrum from a lock-in amplifier measurement with an a.c modulation frequency of 1 kHz is shown in Figure 3.15. The time constant on the
PSD was set at 10 ms which resulted in a corner frequency at 100 Hz as is evident in Figure 3.15.
Nonetheless, the 1/f noise spectrum is clearly visible below 100 Hz and supports the notion that the PSD output contains the conductance fluctuations of the device as we showed in Equation 3.4.
To make sure that we are extracting the same noise performance from the lock-in amplifer as from the DC meaurement setup, we carry out measurements of SNR as a function of gate voltage. The definition of SNR in terms of current fluctuations is given in Equation 2.33 and if we write it in terms of conductance fluctuations, we obtain:
g 1 SNR = √m, G = √ (3.5) SG SVFB
where SVFB is the same quantity as that defined in Equation 2.28. Therefore, SNR (as defined in
Equation 2.33) is an equivalent metric whether one considers current or conductance fluctuations, since it reduces to the voltage source of the fluctuations, namely the noise in the flatband voltage
(SVFB) due to trapping/detrapping events. Figure 3.16 shows a typical plot of the DC SNR (DC
32 -4 10
-6 10 2 -8 /G
G 10 S ~ 1/f -10 10
-12 10 -1 0 1 2 10 10 10 10 Frequency (Hz)
Figure 3.15: Normalized noise power spectra of the conductance of a bioFET device as measured by a lock-in amplifier. The 1/f reference (dashed line) shows that the intrinsic 1/f noise of the DUT is not affected by the choice of the AC modulation frequency.
33 350 12000 SNR gm 300 250 10000 gm (nA/V) ) 200 -1 V 8000 150
SNR ( 100 6000 50 0 4000 -1.8 -1.6 -1.4 -1.2 -1.0 -0.8 Vsg (V)
Figure 3.16: SNR and DC transconductance (gm) as a function of gate voltage in the case of a DC source-drain bias.
source-drain bias voltage, Vds) as a function of gate voltage (Vg) as well as the DC transconductance
(gm) as a function of Vg. SNR can also be extracted in the case of an AC modulation of Vds, from the power spectrum of the conductance fluctuations as shown in Equation 3.5. The SNR and AC transconductance (gm, G) are plotted in Figure 3.17 as a function of Vg. It is evident that the SNR is the same as in the case of the DC measurement, which confirms that the 1/f noise fluctuations of the device are superimposed on the AC fluctuations when making a lock-in measurement and thus appear at the PSD output as highlighted in Equation 3.4. Therefore, the lock-in technique is very useful in reducing the 1/f noise contributions from external circuit elements, but not in reducing the fundamental 1/f noise of the DUT itself. This is in fact why it is possible to use a lock-in amplifier to carry out noise measurements of low-frequency resistance fluctuations[54]. One advantage of the lock-in amplifier setup in sensing experiments however, is the stability of the time traces as seen in Figure 3.18. In the case of a lock-in measurement, a DC source-drain bias is not needed, since the AC modulation is sufficient to extract the conductivity at Vds = 0. This in turn contributes to reduced electrochemical reactions at the source and drain contacts which most probably increases
34 35 12000 SNR gm 30 10000 25 g m,G ) -1 8000 20 (nS/V) V 15
SNR ( 6000 10 4000 5 2000 0 -1.8 -1.6 -1.4 -1.2 -1.0 -0.8 Vsg (V)
Figure 3.17: SNR and AC transconductance (gm, G) as a function of gate voltage in the case of an AC modulation of the source-drain bias. the stability of the sensing experiment as well as impoves the lifetime of the bioFET devices.
3.6 Summary
In this chapter, we covered the fabrication of nanoribbon FET devices, with the main focus of improving reliability and yield and the instrumentation design to carry out sensitive, high- throughput measurements. By adding in a top-gate dielectric layer of silicon oxide (20 nm) and focusing on device widths that are 1 micron or larger, the critical dimension being the thickness
(≈ 40 nm), we are able to improve device uniformity and extend the lifetime of these devices to several weeks in an electrolyte environment, which allows for more reliable sensing data and more complex functionalization schemes. In order to increase the throughput of the measurement, an eight-channel amplifier box was designed combined with software written in Labview to concurrently measure the drain current of 8 devices (Gen.1). A second generation of this multiplexed detection setup (Gen.2) was designed to combine the device interface with the amplification circuit on a single PCB board. The relevant bias voltages were then applied via a NI DAQ card which was also
35
pH 8.13
pH 6.06 pH 6.06
pH 4.06 pH 4.06
pH 2.18 pH 2.18
Figure 3.18: pH calibration curve for a bioFET sensor using the lock-in technique. The AC current response (proportional to the conductance at Vds = 0) of the PSD is plotted as function of time, showing the stability of the time traces when an AC source-drain modulation is used. the interface used to measure 8 channels concurrently. Noise measurements were carried out using the same DAQ card for high sampling frequencies but only a single amplification channel (using a low noise current amplifier, SR570) for low noise current-to-voltage conversion. The different measurement setups were compared against each other for the measurement of a nanoribbon device and it was concluded that the portable sensing setup (Gen. 2) had very similar performance to the low noise, single channel measurement setup provided that (1) the frequency was low (< 10 Hz) and (2) the drain current was larger than 10 nA, which is very promising for the eventual design of a portable and sensitive point-of-care measurement setup. We also carried out measurements using a lock-in amplifer (LIA) setup, which allows one to choose a certain modulation frequency such that the noise pickup from other electronic components is minimized. Our measurements of SNR show that the bioFET device noise dominates for both the LIA and DC measurement setups (single channel measurement) and moreover, the choice of modulation frequency in the lock-in technique does not affect the amplitude of the resistance fluctuations measured at the output of the PSD, resulting in the same SNR for both DC and AC measurements.
36 Chapter 4
Low Frequency Noise of BioFETs
4.1 Introduction
For large area metal-semiconductor-oxide devices, the most common method of probing ox- ide/interface and mobile charges is the use of capacitance voltage measurements[41]. However, for nanoscale field-effect transistor(FET) devices, such measurements are very challenging owing to the small capacitance of these devices, which is easily overwhelmed by parasitic capacitances.
Alternative methods which are both sensitive and non-intrusive so as not to irreversibly alter the electronic properties of the devices are highly sought after. One such method involves the use of noise measurements which attempt to extract information from the dynamic processes involved in the interactions of channel electrons with defects, charge centers and interface states[38][55][56]. One of the major concerns with the down-scaling of electronic devices is the decrease in signal-to-noise ratio since the channel current becomes more prone to fluctuations due to surface/interface states as the surface to volume ratio is increased[57]. Therefore, not only do noise measurements serve as a way to quantify and hopefully reduce the noise fluctuations, but they also allow users to compare across different material systems, device architectures and measurement conditions, in order to identify devices with better noise performance. As researchers began to experiment with nanowire FETs based on different semiconductor materials with different mobilities and gate dielectrics, it became
37 obvious that there was a need for a metric to compare across different device characteristics as well as measurement conditions[46]. Such a metric is Hooge’s parameter, αH , defined in Section 2.4 which is independent of device parameters and measurement conditions[58]. This metric focuses on the noise performance only, but its power lies in the simplicity of its measurement/extraction and the universality of the parameter. Along those lines, a plot of the Hooge’s parameter for various material systems and device architectures will immediately shed light on which combinations result in the best, low-noise performance. Such a plot is shown in Figure 4.1. We can clearly see that devices based on bulk CMOS fabrication have a much lower αH value, on average, than “bottom-up” semiconductor nanowire-based FETs. Figure 4.1 also includes one of our “top-down” silicon-on- insulator (SOI) nanowires (width of 100 nm, height of 40 nm), highlighting the advantage of the top-down process in producing better quality devices with lower noise figures. In this chapter, we consider the use of low frequency noise analysis, namely 1/f noise models, to better understand the physical mechanisms behind current fluctuations in bioFET sensors and guide the fabrication of devices with lower noise figures, with the goal of improving the signal-to-noise ratio and the limit of detection of these devices.
4.2 Effects of Etching Process on BioFET Characteristics
Silicon nanowire (Si NW) and nanometer-scale field effect transistors (FETs) have proven to be quite useful as chemical and biological FETs (bioFETs). However, efforts to fabricate such devices by top- down lithographic methods have often exhibited degraded electrical characteristics associated with the exposed silicon surfaces and high surface-to-volume ratio inherent in NW sensors[10][59][11].
These surfaces have most often been prepared by plasma-etching techniques, long known to cause surface states and bulk damage[60][23]. The presence of these defects has likely been the source of high levels of low frequency noise (LFN), diminished sensitivity, and threshold voltage hysteresis in NW sensors. Anisotropic wet etching has been previously proposed as a method for producing bioFETs with high-quality surfaces as compared to plasma-etched nanostructures. While devices
38 Figure 4.1: Summary of Hooges parameter (αH ) for different nanowire materials as well as sub- micron MOS structures utilizing high-k dielectrics. Included is our best SOI silicon nanowire device (red circle). The dash-dotted line shows the ITRS roadmap specification of αH for the 45nm technology node (Adapted from [46]).
39 200 nm
Figure 4.2: SEM images of two nanowires (NWs) etched using an orientation dependent wet etch, TMAH, and the other one etched using a dry RIE etch chemistry involving Cl2. TMAH etches (100) planes faster than (111) planes and the resulting trapezoidal shape can be clearly seen in the SEM image. The device etched using Cl2 has a more rectangular cross section as expected from an anisotropic RIE etch. have been fabricated using tetramethylammonium hydroxide (TMAH) to produce smoothly faceted
NW devices[23][61], it has not been demonstrated conclusively and quantitatively that the wet- etching process is responsible for the superior electrical performance of these devices. The only semi-quantitative comparison was carried out in our lab using pH sensitivity as the benchmark[11].
It was observed that the RIE etched device showed a lower signal change than what was expected for a TMAH etched device of the same dimensions, for the same change in pH of the buffer solution.
We verified these claims and placed them on more solid, quantitative grounds by comparing nominally identical bioFETs fabricated by TMAH anisotropic etching, as well as two common plasma-etching methods, namely Cl2 and CF4. SEM images of the etched nanowire channels are shown in Figure 4.2. The NW channels for the devices in this investigation were all nominally
100 nm wide and 2.5 µm long. The gate voltage was applied to the back-gate and drain-to-source voltage was kept constant at 0.1V for all measurements. The drain current in this case was measured with a Keithley 2636 source-meter at a sampling frequency of 25 Hz. The equilibrium DC signal was subtracted from the measured current values and the pwelch algorithm was applied to the data using software written in Matlab. The LFN of our devices follows Hooge’s equation given by
Equation 2.27. The frequency exponents for our devices were found to lie in the range 0.8< β <1.5.
1/f β noise is usually characterized by an exponent 0.8 < β < 1.2 [40]. The slightly higher exponents
40 Figure 4.3: Plot of noise amplitude, A, as a function of the drain-to-source voltage (Vds) at a fixed gate voltage, clearly showing that A is independent of the drain current if the number of charge carriers remains unchanged. encountered in our case, as can be seen in Figure 4.4, can be understood by the presence of RTS signals superimposed on the mobility fluctuation 1/f noise[62]. In the present situation, we do not expect to be able to observe RTS on top of 1/f noise since we will show that our devices satisfy the “rule of thumb” N > 1/αH [63]. Exceptions to the rule of thumb are possible in the case of inhomogeneous samples, which would explain additional RTS noise giving rise to a higher 1/f exponent than expected for pure 1/f noise[64][65]. Nonetheless, we confirm the validity of the mobility fluctuation model by observing the invariance of the noise amplitude A = αH /N on the drain current (refer to Figure 4.3), measured at a fixed gate voltage and varying Vds, as well as the linear dependence of 1/A on the gate voltage at fixed Vds (refer to inset of Figure 4.4). Since, in the determination of αH , we operate the devices in the linear region of the transfer curve, the number of carriers is given by: C N = (V − V ) (4.1) e g th where the back-gate capacitance C is determined from simulation results using Altas Device 3D from
Silvaco. The simulation was carried out by a former student[49] and the results of the capacitance voltage simulations were used to extract the value for αH . Therefore, from the slope of the inset
41 Figure 4.4: Typical dependence of 1/f noise spectra on gate voltage for a TMAH-etched device, from which the noise amplitude A, at each gate voltage (13 - 25 V), can be extracted. The inset shows 1/A plotted as a function of Vg, where the slope of the line is used to calculate αH . in Figure 4.4 which is equal to C , the Hooge’s parameter can be determined for each of the e αH
3 types of etched NW devices. The results are shown in Figure 4.5. The TMAH-etched devices have a considerably lower Hooge constant (close to an order of magnitude lower) than the plasma- etched devices and show smaller device-to-device variations. The average Hooge constant for the
−3 TMAH-etched devices is given by αH = 2.1 × 10 , which is comparable to the Hooge constant
−3 values reported for submicrometer MOSFETs with a metal/HfO2 gatestack (αH = 1.6 × 10 for
3 NMOS and αH = 6.9 × 10 for PMOS[66]. The average Hooge parameters for the plasma etched
−2 −2 devices are 1.5 × 10 for Cl2 devices and 1.7 × 10 for CF4 devices. An alternative approach based on a trapping/detrapping model[40][67] to analyze the noise measurement data makes use of the empirical relationship: M g 2 S = m (4.2) I 2 β Cox W L f
42 Figure 4.5: (a) Measured Hooge parameters for three sets of devices. Each set was etched using either TMAH or Cl2 or CF4. The box plot shows the 25th percentile, the median, and the 75th percentile (the mean is indicated by asquare marker). The average values of αH were 0.0021 for the TMAH devices, 0.015 for the Cl2 devices, and 0.017 for the CF4 etched devices. (b) Measured subthreshold swing for three sets of devices, etched using either TMAH or Cl2 or CF4. The average value for the TMAH devices was 1.0 V/decade. For Cl2 etched devices, the average was 2.6 V/decade, and for CF4 devices, the average was 3.0 V/decade. where M is the parameter to be extracted. W and L are the width and length of the channel, respectively, gm is the transconductance at the operating point of the measurement, and Cox is the gate capacitance per unit area. Using the aforementioned model, the average M parameter for the TMAH etched devices is M = (6 ± 2) × 10−24. Likewise, the average values for the plasma-
−24 −24 etched devices are M = (22 ± 10) × 10 for the Cl2 devices and M = (12 ± 4) × 10 for the CF4 devices. A similar trend is observed for the noise parameter M, where the TMAH-etched devices exhibit a lower noise figure than the plasma-etched ones.
The subthreshold swing measurements for each type of device are also shown in Figure 4.5(b).
Subthreshold swing, SS, is given by the following:
k T C + C + C SS = 2.3 ox d it (4.3) q Cox
43 where Cd is the depletion capacitance per unit area of the silicon channel and Cit is the capacitance per unit area associated with interface traps. The measured values of S are all quite high due to the thickness of the buried oxide (405 nm) serving as the gate insulator, which makes Cox small. However, the TMAH devices exhibit significantly lower average subthreshold swing and smaller device-to-device variation than either of the plasma-etched devices, which follows the trend exhibited by the values for αH. The differences in SS are attributed to the different values of Cit in the case of TMAH devices versus plasma etched devices. We attribute the lower noise figure and lower subthreshold swing, in the case of the wet-etched devices, to a lower density of surface states at the etched sidewalls. The differences in oxide capacitance and nanowire channel depletion capacitance are negligible and cannot account for the differences in the measured subthreshold swing, which we ascribe to variations in the interface state density. The noise and subthreshold swing measurements, therefore, together quantitatively confirm that wet-etching based methods yield smoother surfaces and consequently better electrical characteristics.
4.3 Temperature Dependence of 1/f Noise Mechanisms
The analysis of the previous section was focused on the linear regime of operation, which is where Hooge’s equation is applicable (see Equation 2.27). For a more complete modeling of the noise, from subthreshold to strong accumulation or strong inversion, the number fluctuation model
(Equation 2.29) and the correlated number-mobility fluctuation model (Equation 2.30) are more relevant. To fully characterize the 1/f noise of our Si NWs, we carried out measurements from weak to strong inversion at low drain bias and at different temperatures (100 to 300 K). We observe a change in the noise mechanism as the temperature is lowered, from the ∆n-∆µ correlated model to the ∆n model. At certain gate voltages, the 1/f spectra evolve into Lorentzian spectra, indicative of random telegraph signal (RTS) noise. The NW channels were defined by a wet orientation dependent etch using tetramethylammonium hydroxide (TMAH), which we have shown yields better electrical characteristics and lower noise figures compared to dry etching techniques. The devices are used
44 Figure 4.6: Typical noise spectra measured for a device at different gate voltages (22 - 40 V) from subthreshold to strong inversion. The exponents of the 1/ f spectra all lie in the range 0.8 < β < 1.2. The flattening of the noise spectra is due to background noise from the measurement setup.
in the back-gated configuration, where the gate voltage Vg is applied to the handle layer of the
SOI wafer. The source-drain bias was kept constant at 0.1 V and the drain current Id was fed into a current preamplifier (SRS 570), setup as a band-pass filter with cutoff frequencies at 0.03 and
1 kHz. The noise data was recorded using a National Instruments DAQ card at 1000 samples/s and the spectral power density estimated using the pwelch function in Matlab. The temperature dependent measurements were carried out in a Janis Instruments variable temperature cryostat.
With the DAQ card measurement setup, noise spectra were measured up to a frequency of 500 Hz, and typical spectra are shown in Figure 4.6.
2 The normalized drain current noise, SI/I at f=1 Hz (noise amplitude, A), is extracted and plotted
2 against Id in Figure 4.7. In the subthreshold region at 300 K, it is observed that SI/I is proportional
2 to (gm/I) , where gm is the transconductance extracted from the Id −Vg curve. Since the correlation does not hold as well in the strong inversion region, we conclude that the 1/f noise of our devices
2 follows the ∆n-∆µ correlated model. For measurements at 100 K, the SI/I versus Id plot indicates
2 a change in the dominant 1/f noise generating mechanism, since SI/I is now proportional to
45 Figure 4.7: Normalized drain current noise amplitude at f=1 Hz (A) is plotted against drain current 2 (Id). Measurements at room temperature (300 K) compared to the gm/I curve clearly indicate that the device follows the correlated ∆n-∆µ noise model. For measurements carried out at 100 K, the noise amplitude is consistent with a carrier number fluctuation noise ( ∆n) model as indicated by the change in slope compared to the 300 K data in the strong inversion region.
2 2 2 (gm/I) even in strong inversion. The enhanced correlation between SI/I and (gm/I) indicates that at low temperature the correlated mobility fluctuations are suppressed and carrier number
fluctuations dominate the LFN[68]. The transition from the ∆n-∆µ correlated model to a purely
∆n model is a gradual transition as seen from the change in slope in strong inversion in Figure 4.8
(a). The data for T=250 K and T=100 K have been scaled to make the change in slope clearer, since the noise is subthreshold increases with decreasing temperature. Even though only the data for T=100 K is shown, the measurements carried out at T=200 K are proportional and show the same trend as the T=100 K data. This means that the transition happens over a small temperature range, between 200 K and 300 K. For the carrier number fluctuation model, including correlated mobility fluctuations, the gate-referred noise is given by Equation 2.30. For negligible correlated mobility fluctuations we expect Svg to be equal to the flatband-voltage noise density (SVFB), but for appreciable values of the scattering coefficient α and large number of carriers (i.e operating in
46 Figure 4.8: (a) The data show the gradual change of the slope in the strong inversion regime as temperature is lowered at T=300, 250, and 100 K, respectively. The data points for 250 and 100 K have been scaled for easier visualization of the change. (b) The data are fitted using the correlated ∆n-∆µ noise model, confirming the change in the noise generating mechanism as the temperature decreases. The Coulomb scattering coefficient, α, is also extracted from linear fits to the data at 250 and 300 K. The fit to the data at 100 K clearly indicates that the correlated model is no longer valid at that temperature.
47 strong inversion), we expect a linear relationship between Svg and Vg. We therefore investigated how well the data in Figure 4.8(a) fits the correlated ∆n-∆µ model as temperature is lowered. For the measurement at 300 K in Figure 4.8(b), the data confirm what we expect, with Svg directly proportional to Vg. As the temperature is lowered to 250 K, the linear relationship becomes hard to resolve due to the scatter in the data points, but a linear least-squares fit gives an estimate of the dependence. For 100 K, the absolute noise magnitude is much larger as expected from previous observations and the direct proportionality between Svg and Vg vanishes. Instead, the gate-referred noise increases as gate voltage is decreased. In this regime, the correlated mobility fluctuations have been completely suppressed and what we observe is the gate voltage dependence of the carrier number fluctuation noise, or in other words the gate voltage dependence of the interface trap density. √ The calculated slopes correspond to α Cox µ SVFB. Both the mobility and the flatband-voltage noise density have a slight temperature dependence from 300 K down to 250 K. Accounting for this, it is determined that the scattering coefficient α decreases as the temperature is lowered, from
1.9 × 104 V · s/C (at 300 K) to 96 V · s/C (at 250 K). The capacitance values were obtained from
Atlas 3D simulations as mentioned in the previous section. As temperature is lowered, it is also observed that mobility increases due to reduced phonon scattering. This trend is observed upto the lowest temperature measured (around 100K). Lower temperatures were impractical due to the fact that the handle layer was not degenerately doped and therefore at low temperatures, freeze-out of the dopant atoms render the handle layer completely insulating and back-gating becomes impossible.
The temperature dependent measurements of mobility for two devices are shown in Figure 4.9.
The mobility is calculated from the measurement of the peak transconductance (extracted from numerical differentiation of the Id-Vg curve), through the following:
L g µ = m (4.4) W Cox Vds
Thus, both mobility scattering and mobility fluctuation noise is suppressed in those devices as temperature is lowered.
48 Figure 4.9: Plot of mobility as a function of temperature for 2 NW devices, showing the increase in mobility as temperature is decreased due to suppressed phonon scattering.
For devices where a large number of trapping/detrapping events are involved, the superposition of multiple Lorentzian spectra (signature of a two level random telegraph signal) results in 1/f- type spectra[37]. However, for certain combinations of gate bias and temperature, only a few traps are involved in the carrier number fluctuation and we are able to decompose the 1/f spectra into its constituent Lorentzian components[69][70]. Figure 4.10 explicitly shows the evolution of the 1/f spectrum at room temperature to a Lorentzian curve superimposed on a 1/f spectrum, as temperature is lowered to 120 K. The primary effect of the decrease in temperature is to reduce the number of traps which are thermally accessible, resulting in a noise spectrum originating from a few
fluctuators only. 1/f noise is still present for frequency values much less than 25 Hz, but the roll-off at approximately 25 Hz and the subsequent 1/f2 dependence clearly indicates a superimposed RTS power spectrum. We will investigate RTS spectra further in the next section. This decomposition of the 1/f noise into a regime where a few active traps dominate the noise power spectrum provides very strong evidence for trapping/detrapping processes as being the physical origin of 1/f noise[37].
4.4 Random Telegraph Signals
Random Telegraph Signals (RTS) are discrete transitions in the current level which represent the interactions of electrons with a single trap, of a well defined energy level. A simple schematic of such
49 Figure 4.10: Evolution of the normalized noise power spectral density from room temperature (300 K) to 120 K for a NW device. At certain gate voltages and at low temperature (120 K), only a few traps are active and contribute to the noise spectrum, which consequently changes from a 1/f spectrum (with the dotted line representing a 1/f least-squares fit) to a Lorentzian superimposed on a 1/f trend, evident for the larger frequencies (with the solid line representing a Lorentzian least-squares fit with a corner frequency of 25 Hz). signals is shown in Figure 4.11. The power spectrum of RTS can be modeled through a Lorentzian as such: K Γ S = (4.5) f Γ2 + (2 π f)2 where K is usually a fitting parameter describing the amplitude of the RTS, Γ is a function of the two characteristic time constants, τ0 and τ1, representing the average time spent in the low and high current states respectively. Γ can be calculated from:
1 1 Γ = + (4.6) τ0 τ1
In the previous section, we showed how 1/f noise spectra can be decomposed into Lorentzian spectra (characteristic of RTS). Figure 4.12 shows how multiple Lorentzian spectra, from discrete single level trapping events, can be combined to give a 1/f spectrum, within a certain frequency range. For certain combinations of temperature and gate voltage, we observed clear RTS signals
50 I (A)
τ1
τ0
Time (s)
Figure 4.11: Illustration of RTS in a current-time trace, showing two discrete current levels with characteristic times τ0 and τ1, representing the low and high current states respectively.
Figure 4.12: Computed results showing how the addition of Lorentzian spectra due to discrete two-level trapping systems can and indeed does result in 1/f noise spectra for an ensemble of such traps.
51
Figure 4.13: A segment of drain current versus time measurement for a NW bioFET device showing the discrete two-level switching, indicating the activity of a single trap resulting in RTS. in our NW devices as can be seen from the time trace in Figure 4.13. The noise power spectra for different gate voltages are shown in Figure 4.14 and it is obvious that the shape follows the
Lorentzian spectrum given in Equation 4.5. As the gate voltage is increased to 27 V, it is observed that the Lorentzian spectrum starts to change to a more 1/f type spectrum.
To obtain more information about the nature of the trap, we investigated the mean times spent in the low and high current states (τ0 and τ1 respectively) and their dependence on gate voltage first and secondly on temperature. The histograms in Figure 4.15 show that as gate voltage is increased, the current spends more time in the “low” state. We associate the “high” current state with a neutral trap state, since that is the condition under which coulombic scattering is minimized. As the gate voltage is increased, the Fermi level gets closer to the conduction band (for an n-channel device in this case) and the probability of a trap being occupied increases. Therefore, we conclude that the “low” current state is associated with a filled trap state and the “high” current state is associated with an empty trap. Since the trap is charged when filled and neutral when empty, we conclude that the trap is an acceptor type trap[71]. It is common and more informative to refer to the characteristic emission and capture times (τe amd τc) instead of τ0 and τ1, when characterizing
52
Figure 4.14: Normalized noise spectra at different gate voltages showing the typical Lorentzian spectra associated with RTS. As the voltage is increased to 27 V, it is clearly seen that the Lorentzian spectrum is changing to a 1/f spectrum, indicating a gate voltage dependence of the RTS.
a certain trap level. The characteristic emission time (τe) refers to the average time it takes before an electron is emitted from the trap whereas the capture time refers to the average time is takes before an electron is captured into the trap. After identifying the nature of the trap (whether it is an acceptor or donor), it is straightforward to determine which of τ0 or τ1 corresponds to τe. In our case, since the trap is an acceptor, the “low” or “0” current state is associated with a filled trap and therefore, τ0 = τe and τ1 = τc. For each current time trace (such as that shown in Figure 4.13) we assign a certain threshold current value to differentiate between the two states, 0 and 1. The time spent in each state was then determined and a Poisson fit was applied to the distribution of capture and emission times to extract the characteristic τe and τc for each RTS time trace. Figure 4.16 shows the extracted characteristic times τe and τc at different gate voltages. The inset shows an example distribution of emission times along with the exponential Poisson fit which allows us to extract a characteristic emission time. From the plot in Figure 4.16, we confirm our deduction regarding the nature of the trap since τe increases and τc decreases as the gate voltage is increases, which means that the electron spends more time before being emitted as the trap occupancy increases.
53 We can further model the two-level switching as a thermally activated process between two states, where the average time (τi) spent in state i is given by an Arrhenius law[70]:
E τ = τ exp i (4.7) i 0,i k T
where Ei is the activation energy required to leave state i and 1/τ0,i corresponds to an attempt frequency which is usually on the order of the inverse phonon frequency. By varying the temperature, we are able to plot the characteristic emission and capture times in Figure 4.17. On a semilog plot, the slope allows us to determine the activation energy of the capture process (Ea, c
= Ea, 1 = 250 meV) as well as the activation energy of the emission process (Ea, e = Ea, 0 = 180 meV). The emission and capture times can be used to independently calculate the corner frequency
(fc) using Equation 4.6 where it is important to note that Γ = 2 π fc. fc can also be estimated directly from the Lorentzian spectra in the frequency domain (see Figure 4.14). The estimated corner frequencies are in good agreement with the calculated ones using Equation 4.6. It is found that fc also follows an Arrhenius law, with an activation energy of 230 meV (see Figure 4.18).
By considering the statistics of capture and emission involving the occupancy of a single trapping level, the following relationship between the capture and emission times can be derived[72]:
τ E − E c = g exp T F (4.8) τe k T
where (ET − EF) is the trap energy level relative to the Fermi level and g is the degeneracy factor which is usually considered to be one. Combining Equations 4.8 and 4.7, we deduce that:
τ0,c ET − EF = Ea,c − Ea,e + kT ln (4.9) τ0,e
The term kT ln τ0,c was calculated to be equal to 60 meV, which results in a trap level that is 10 τ0,e meV below the Fermi level. Since the measurements of the RTS were carried out in the subthreshold
54 Vg = 25 V
Vg = 26 V
Vg = 27 V
Figure 4.15: Histogram of current values showing the bimodal distribution characteristic of a two- level switching signal. The changes in trap occupancy as the gate voltage is varied can be seen as changes in the populations of the “low” and “high” current states. From the changes evident in the histograms, we conclude that the active trap in this case is an acceptor trap, that is the trap is charged when filled and neutral when empty.
55 T = 130K 90 2.8 80 -t/τ 70 ̴ e 2.6 60 2.4 τ c
50 (ms) 40 2.2 (ms) e
τ 30 2.0 20 10 1.8 24 25 26 27 Vg (volts)
Figure 4.16: Plot of the capture (τc) and emission (τe) times as a function of gate voltage at a temperature of 130 K. The inset shows the typical distribution of times which follow a Poisson distribution. An exponential or Poisson fit allows us to extract the characteristic emission and capture times.
Emission Capture
Figure 4.17: Arrhenius plot of the characteristic emission time (τ0) and capture time (τ1) as a function of temperature. From the slope, the thermal activation energy of the capture and emission processes can be extracted.Ea, e = 250 meV and Ea, c = 180 meV.
56 Figure 4.18: Arrhenius plot of the corner freqency (estimated from the noise power spectra) at different temperatures. The linear fit shows that fc can be modeled by a thermally activated process with an activation energy of 230 meV. regime of the FET device, we assume that the Fermi level is close to mid-gap. This results in an acceptor trap level that is about 0.57 eV from the conduction band edge, which is consistent with deep acceptor levels that are caused by the presence of gold[73]. In our case, the source and drain contacts were formed from evaporated titanium followed by gold which might have migrated closer to the channel during the annealing step.
The relative amplitude of RTS can be modeled as[74]:
∆I 1 1 d = ± αµ (4.10) Id WL N where α is the scattering coefficient we encountered earlier in Equation 2.30 and N is the number of carriers per unit area. These devices show extremely large relative amplitudes (50 - 120%) compared to the values that have been reported in the literature[75]. The relative amplitude as a function of 1/Id is plotted in Figure 4.19. As we expect, the relative noise amplitude is linearly proportional to the inverse of the drain current (Id) with the slope of the line being related to the device dimensions as well as the effective field-effect mobility (µeff ). Since the slope does not change much in the temperature range investigated, we conclude that the changes in mobility are
57
1.2
1.0
0.8 ∆Ι/Ι
0.6 110K 120K 0.4 130K
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 9 -1 1/Id (x10 A )
Figure 4.19: Relative amplitude of RTS plotted as a function of the inverse of drain current (1/Id). The linear fit shows that the data agrees well with the 1/N dependence of relative RTS amplitude. The large changes in the level of the RTS noise is attributed to changes in the scattering coefficient α, resulting in different levels of mobility scattering. negligible. Therefore, any changes in the intercept are due to changes in the scattering coefficient α.
As temperature is lowered, the value of α is found to increase, resulting in more mobility scattering and thus a larger relative amplitude of the RTS. These large amplitudes of RTS present a major problem for ultra-scaled devices where the effects of a few scattering centers become much more significant. Therefore, even though gold is an inert material for electrochemical and biochemical applications, its use in the fabrication of bioFETs should be avoided for better noise performance.
4.5 Summary
In this Chapter, we have sought to apply noise analysis techniques to silicon nanowire bioFETs as a non-destructive and accurate tool to characterize the noise performance of such devices, with the goal of improving the signal-to-noise ratio. We have demonstrated the use of anisotropic wet orientation dependent etching with TMAH as a method of producing bioFETs with high-quality surfaces and, consequently, superior electrical characteristics as compared to plasma etched surfaces.
Our results lead us to conclude that, as bioFET devices are downscaled further and the surface-to- volume ratio increases, the particular etch process used will be critical in determining the density of
58 surface states and, ultimately, the noise performance. We have also characterized the temperature dependence of the LFN behavior of NW bioFETs. At room temperature the devices are shown to follow the correlated number-mobility fluctuation model with noise spectra that show a 1/f dependence even in subthreshold. As the temperature is lowered, we observe a suppression of the correlated mobility fluctuations, yielding a pure carrier number fluctuation noise. For even lower temperatures, the LFN noise is decomposed further to a few carrier number fluctuators, which give rise to a Lorentzian power spectrum instead of a 1/f spectrum. For certain combinations of gate voltage and temperature, we were able to probe the regimes where we would observe very strong RTS time traces (with relative amplitudes >100%). Using time-domain as well as frequency domain analysis techniques, we were able to extract the characteristic times and activation energies for capture and emission and also determined that the trap was due to a deep acceptor level 0.57 eV from the conduction band edge, indicative of the presence of gold. The measurements carried out in this Chapter clearly indicate that noise fluctuations can be a very significant issue for highly scaled devices and to ensure that such miniaturization can be pursued, while keeping the signal-to-noise ratio unchanged, will require a deep understanding of the noise generating mechanisms in such devices. For instance, the RTS analysis demonstrating the presence of gold clearly shows that the use of gold is a significant source of current fluctuations and therefore should be avoided in the design of bioFET sensors even though gold is a great inert material for a lot of biochemical/electrochemical applications. Furthermore, noise analysis can be used as a benchmarking tool to identify process steps that yield a lower noise figure for instance and thus, guide the fabrication and development of bioFET sensors with lower noise and consequently lower LOD.
59 Chapter 5
Signal-to-Noise Ratio as a
Performance Metric
5.1 Introduction
Noise analysis is a tool that has only recently been adopted by the nanowire research community as described in Chapter4 and only more recently been applied to the field nanowire/nanoribbon based biosensors[36][76][35] with the realization that the noise properties will drastically impact the smallest signal changes that can be measured as the detection limits are being pushed further and further. A commonly used metric in the field of biosensors is the sensitivity, defined as the relative change in the signal (∆I/I), which allows for comparisons across devices with different dimensions and transfer characteristics. Consequently, a lot of theoretical as well as experimental conclusions are based on the optimization of this performance metric[13][32]. However, as far as the limit of detection of sensors is concerned, this metric fails to account for the primary limitation which is the noise of the particular sensor system/device. As devices are scaled down to the nanometer regime, the channel current becomes more prone to fluctuations due to oxide traps and interface states and the low-frequency noise of the device becomes a very serious limitation for any DC
60 measurement as we have shown in Chapter4. Therefore, it becomes essential, in the design of better sensors with low limits of detection (LOD), to consider a performance metric which includes the signal transduction/amplification as well as the effects of noise fluctuations, in order to provide a complete physical model to understand the parameters which affect the LOD. Signal-to-noise ratio
(SNR) can be used as such a metric[77] since it involves both the device transconductance (gm) which is directly proportional to the signal generated as well as the current noise power density
(SI ) which can be modeled using the correlated ∆n-∆µ model given in Equation 2.30 for instance.
An equation for SNR was derived in Chapter 2 (see Equation 2.33) with the very simple result that the LOD (which is given by 1/SNR) is roughly equal to the amplitude of the flatband voltage noise.
Therefore, using the noise measurement tools described in Chapter4, SNR can be extracted for any device, prior to sensing experiments and consequently be used to either compare devices with different LOD or screen the devices for the lowest LOD.
5.2 Sensitivity
The sensitivity of FET biosensors is usually defined as ∆I/I. This relative signal change, which is a property of the device and measurement conditions, needs to be distinguished from the sensitivity of the chemical/biological transduction layer, which turns out to be more relevant for biosensing applications. For example, when considering the pH sensitivity of a device or surface, the value is usually reported as a change in device signal per unit change in the measurand, in this case pH value of the solution. The relative signal change is a function of the bias point (Vg − Vth) as can be seen in Equation 5.1. Looking at the latter equation, it is obvious that the closer the gate voltage
(Vg) is to the threshold voltage (Vth), the larger the relative signal change (∆I/I). However, as can be seen in Figure 5.1, the larger the relative signal change, the larger the relative noise level as well. Figure 5.1 shows the same pH measurement (from pH 7.1 to 7.9) carried out at different gate voltage bias points, resulting in different relative signal changes as given in Equation 5.1. As mentioned earlier, using the relative signal change as a performance metric ignores the fact that
61 the amplitude of the noise also varies with the sensitivity.
∆I ∆ψ = 0 (5.1) I (Vg − Vth)
The true sensitivity remains unchanged as can be seen from Figure 5.2. The device response (∆I) scales linearly with gm, as expected. From the linear fit, the change in surface potential for a unit change in pH can be extracted as 39 mV/pH. This clearly indicates that the true sensitivity is independent of the bias point that is chosen. It only affects how the surface potential change is extracted from the current signal change. Figure 5.2 also highlights the fact that the measurement itself is a measurement of ∆I and since any bias point (i.e. gm values) can be used to extract the true surface sensitivity, we should focus our attention on where the device noise is minimized so that the measurement is carried out with the highest signal-to-noise ratio (SNRmeas) possible.
A similar investigation was carried out for biotin-streptavidin binding. This model system is often times used for its strong binding affininity (∼ 10−14 M)[78], resulting in very clear and stable signals from the binding events. The functionalization is also more robust and straightforward for this binding system which results in more consistent and comparable sensing data. Streptavidin, being a tetrameric protein, consists of four binding pockets which can potentially bind to four biotin molecules. In the case of surfaces functionalized with biotin, the bound streptravidin molecules still contain unconjugated binding pockets which can interact with other biotin molecules in the solution bulk and therefore can be competitively desorbed if the bulk concentration is much higher than the surface concentration. This process is schematically shown in Figure 5.3. Using such a competitive desorption process allows us to re-use the same bioFET device, yielding results that are easier to compare and thus quantify. The bioFET response for streptavidin binding followed by competitive desorption using D-biotin is shown in Figure 5.4. At the pH used in this experiment (pH 7.5) the charge on each streptavidin molecule is about −2 e [79], which results in a decrease of the drain current on streptavidin binding (The devices used for this investigation are n-channel FETs). The complete restoration of the baseline current provides strong evidence that the competitive desorption
62 (a)
(b)
Figure 5.1: (a) Normalized current data for a pH change from pH 7.1 to pH 7.9 at different bias points (different Id levels). (b) Corresponding Id-Vg curve, where the bias points used in (a) are indicated with crosses. For measurements done closer to the threshold voltage, the resulting ∆I/I is larger as can be seen from Equation 5.1 and the experimental data in (a). The relative noise level is also larger.
63 Figure 5.2: Plot of device reponse (∆I) as a function of the transconductance (gm at different gate voltage bias) for a pH change from 7.1 to 7.9. From the linear least squares fit, the pH sensitivity of the device was extracted as 39 mV/pH. strategy works and therefore can be used to regenerate surfaces with bound streptavidin. Using this regeneration method, we carried out detection of the same concentration (20 pM) of streptavidin at different bias points, resulting in different normalized current changes (different ∆I/I) as shown in Figure 5.5. As expected, the relative signal change is larger for gate voltage values closer to the threshold voltage (see Equation 5.1). In the normalized current response plot, it is also seen that the relative noise amplitude is larger for gate voltage values closer to the threshold voltage.
The signal-to-noise ratio for this measurement can be extracted from the ratio of the response to
D-Biotin D-Biotin
Figure 5.3: Illustration of the competitive desorption process of surface bound streptavidin molecules using a high bulk concentration of D-biotin molecules. The surface bound streptavidin consists of unbound sites that are available to molecules in the bulk solution. By introducing D- biotin to the bulk, the strength of that interaction is enough to pull the streptavidin molecules away from the surface bound biotin molecules.
64 Manual Switching
D-Biotin
Streptavidin
Figure 5.4: BioFET response to streptavidin binding (20 pM) in 0.01X PBS, followed by competitive desorption using 1mM D-biotin. The initial current drop is due to the positive charge of streptavidin and the subsequent restoration of the baseline current, on addition of D-biotin, shows that the competitive desorption strategy works. The sampling rate used was 1000 Hz.