Bm

JFa

b

c

a

ARRAA

KMEBE

1

ttnsmdpmPuTl

(

0h

Sensors and Actuators B 192 (2014) 283– 293

Contents lists available at ScienceDirect

Sensors and Actuators B: Chemical

journa l h om epage: www.elsev ier .com/ locat e/snb

ayesian decoding of the ammonia response of a zirconia-basedixed-potential sensor in the presence of hydrocarbon interference

ulia Tsitrona, Cortney R. Krellerb, Praveen K. Sekharc, Rangachary Mukundanb,ernando H. Garzonb, Eric L. Broshab,∗, Alexandre V. Morozova,∗∗

Department of Physics & Astronomy and BioMaPS Institute for Quantitative Biology, Rutgers University, Piscataway, NJ, USALos Alamos National Laboratory, Los Alamos, NM, USAWashington State University Vancouver, Vancouver, WA, USA

r t i c l e i n f o

rticle history:eceived 24 July 2013eceived in revised form 25 October 2013ccepted 27 October 2013vailable online 5 November 2013

eywords:ixed-potential sensor

lectrochemical sensorayesian modelingngine exhaust analysis

a b s t r a c t

Zirconia-based mixed-potential sensors are a promising technology for monitoring levels of nitrogenoxides and ammonia in diesel engine exhaust. However, in addition to the target gases these sensors reactto unburned hydrocarbons present in the gas mixture. The observed cross-interference between targetand non-target gases cannot be fully mitigated by applying different bias currents to the sensor. On theother hand, sensor sensitivity and selectivity toward various components of the mixture depend on thebias current setting, allowing us to effectively create an array of sensors by applying different bias currentsto the same device. Here we show how such an array can be used to predict absolute concentrations ofammonia in the presence of propylene. Our Bayesian framework can be easily generalized to other typesof sensors and to more complex chemical mixtures. It consists of two steps: the calibration step, in whichthe parameters of the model are determined a priori in the laboratory setting, and the prediction step,

which mimics the deployment of the device in real-world conditions. We investigate a linear model, inwhich response of the sensor to each gas is assumed to be additive, and a nonlinear model, which takesinterference between gases into account. We find that the nonlinear model, although more complex,yields more accurate predictions. We also find that relatively few sensor readings and bias current settingsare required to make reliable predictions of gas concentrations in the mixture, making our approachfeasible in a variety of automotive and other technological settings.

. Introduction

The potential health consequences resulting from exposureo diesel engine exhaust have been well documented; theseransportation-related pollutants are responsible for a number ofegative health and environmental impacts [1,2]. However, newtudies have shown that recently strengthened federal require-ents governing diesel engines of new manufactured heavy-duty

iesel vehicles have resulted in major cuts in emissions ofarticulate matter (PM) and nitrogen oxides (NOx) [3]. After imple-entation of new standards in the US in 2010, the NOx and

M emissions from the diesel vehicles studied had been reduced

pwards of 98% and 95% respectively per gallon of diesel fuel [3].hese reductions stem from the use of selective reduction cata-yst (SRC) [4,5] and exhaust gas recirculation (EGR) [6,7] systems

∗ Corresponding author.∗∗ Corresponding author. Tel.: +1 8484451387.

E-mail addresses: brosha@lanl.gov (E.L. Brosha), morozov@physics.rutgers.eduA.V. Morozov).

925-4005/$ – see front matter © 2013 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.snb.2013.10.115

© 2013 Elsevier B.V. All rights reserved.

paired with regenerative particulate traps [8] located in the exhaustsystems of heavy-duty diesel trucks. With SRC type systems, anammonia source is provided by the on-board storage of dieselexhaust fluid, typically a mixture of urea and de-ionized water. SRCsystems also incorporate limited EGR.

In contrast, the enhanced-EGR (EEGR) approach uses exten-sive exhaust gas recirculation in order to deprive the combustionevent of oxygen by introducing cooled exhaust gas into the intakesystem and therefore does not use, or require, urea for on-boardammonia generation [9]. Both systems rely on particulate trapsto reduce PM emissions; the traps are regenerated periodicallyon-board by injection of diesel fuel at elevated temperatures. Theinclusion of PM filters and traps made necessary the transitionto ultra-low sulfur levels in diesel fuels to prevent excessive PMformation and premature clogging. The EGR approach to reduc-ing NOx emissions has a greater negative impact on overall dieselfuel efficiency since more particulates are produced by design

in order to reduce NOx formation within the combustion cham-ber and thus the PM traps must be regenerated more frequently.SRC systems, on the other hand, may offset added weight andextra system costs and complexity by offering greater fuel savings.

2 Actuators B 192 (2014) 283– 293

Td

aatSiceatltesofwaim

smciprrNemrrratbp[Doatnmc

ordfadct(cec(ttog

Fig. 1. Au/YSZ/Pt mixed-potential sensor. (a) Schematic illustration of the sensor.

84 J. Tsitron et al. / Sensors and

he exact approach used by vehicle manufacturers is left to theiriscretion.

Despite these recent advances, there remains a demand for suit-ble exhaust gas sensor technologies to monitor tailpipe emissionsnd to control and maintain efficient operation of SRC and EGR sys-ems [10]. For example, uncontrolled ammonia injection into theRC can cause more serious air pollution than excess NOx result-ng from the combustion event [11]. Ideally, implementation oflosed-loop systems that keep ammonia use close to stoichiom-try in SRC systems in order to monitor tailpipe-out NOx emissionsnd insure regulatory compliance is desired [10–13]. However,he development of successful sensors for diesel applications hasagged in comparison to the advances made in the implementa-ion of oxygen lambda sensors for on-board diagnostics (OBD-II) ofmission systems [14] that have become mandatory for gasoline,park-ignition vehicles. Because of the successful implementationf robust, zirconia-based mixed-potential electrochemical sensorsor gasoline engine control and OBD in conjunction with three-ay catalytic converters, derivatives of this technology for diesel

pplications naturally attract considerable interest. However, themplementation of the diesel analogue to OBD has proved to be

ore complicated.Mixed-potential sensors are electrochemical devices that mea-

ure the non-Nernstian potential of a mixture of gases, where theixed potential is fixed by the rates of different electrochemi-

al reactions occurring simultaneously at an electrode/electrolytenterface [15,16]. The employment of electrode materials thatossess varying electro-catalytic activities toward the redox halfeactions has been shown to further increase the mixed-potentialesponse. Leveraging the success of commercial zirconia sensors,Ox [17,18] and NH3 [19,20] sensors based on non-equilibriumlectrochemical principles have been investigated; however, com-ercialization of this technology has been hindered by its poor

eproducibility and stability [21,22]. Because the mixed-potentialesponse is highly dependent on the kinetics of electrochemicaleactions occurring at each electrode, stable electro-active areasre required to maintain a constant sensor response over the life-ime of the device. Additionally, device sensitivity is improvedy minimizing the amount of heterogeneous reactions occurringrior to the analyte gas reaching the electrochemical interface23,24]. To address these issues, the Electrochemical Sensors andevices Group at Los Alamos National Laboratory (LANL) has devel-ped a patented sensor design that incorporates dense electrodesnd porous electrolytes in bulk [24–26], thin film [27,28], andape cast forms [29] (Fig. 1). By using dense electrodes, heteroge-eous catalysis is reduced, increasing sensitivity, and the increasedorphological stability of dense electrodes yields a robust electro-

hemical interface, increasing lifetime durability.In general, the scientific community has addressed only some

f the various issues with electrochemical sensors such as accu-acy, temperature-dependent sensitivity, response time, sensorrift (baseline and signal), flow-rate dependence, poisoning fromuel constituents, cross-sensitivity, and thermal cycling durabilitynd shock. Cross-sensitivity to non-target chemical species within aiesel emission system is an area of concern that this work specifi-ally focuses on. Existing commercial solid-state electrochemicalechnology (Au and Pt electrodes and yttria-stabilized zirconiaYSZ) electrolyte) exhibits significant cross-sensitivity to hydro-arbons (HCs) when exposed to a complex gas mixture of dieselngine exhaust [30]. In this study, we investigate how the appli-ation of small bias currents to Au/YSZ/Pt mixed-potential sensorsFig. 1) may be used to address this issue. Prior studies have shown

hat sensor selectivity and sensitivity to a given analyte may beuned by using current biasing [28,31–33]. Although the responsef the Au/YSZ/Pt sensor used in this study toward both the tar-et gas, NH3, and HCs is strongly influenced by current biasing, we

The sensor uses wire electrodes embedded into tape-cast, porous YSZ solid elec-trolyte. Top: side view, bottom: view from below. (b) Photograph of the actualdevice. Electrolyte body is 1 cm in length.

find that cross-interference effects cannot be entirely eliminatedwith this method. Here we focus on detecting NH3 in the presenceof propylene (C3H6) as an HC interference. Since the sensor reactsto both interfering and target gases regardless of the magnitudeof the bias current, the presence and concentration of NH3 in thegas mixture cannot be straightforwardly deduced from the sen-sor output. Thus we employ Bayesian inference techniques [34] toanalyze sensor readouts. In order to achieve robust predictions, weuse an array of sensors, each under a different bias current and thusexhibiting differential sensitivity to NH3 and C3H6. The array-basedapproach is reminiscent of mammalian and insect olfactory sys-tems in which, as a rule, several kinds of olfactory receptors respondto a given analyte [35,36]. Similar principles have been employed tocreate cross-reactive synthetic sensor arrays, whose output is typ-ically analyzed using pattern-recognition algorithms [37,38]. Wehave previously developed a Bayesian framework which explic-itly takes receptor–ligand interactions into account, and appliedit to infer concentrations of four highly related sugar nucleotidesfrom the output of four bioengineered G-protein-coupled recep-tors [39]. Here we apply a similar methodology to deduce NH3 andC3H6 concentrations simultaneously, using readouts from an arrayof mixed-potential sensors as input. We also study the accuracyof our predictions as a function of the number of sensors in thearray and the number of independent measurements. This allowsus to deduce the minimal set of requirements for deployment ofour sensor system in real-world conditions.

2. Materials and methods

2.1. Preparation of sensors

Mixed-potential sensors based on 8 mol% yttria stabilized zir-conia (Tosoh TZ-8YS, Japan) and dense, metal wire electrodes (Au,Pt) with preferential selectivity to ammonia were fabricated using

a tape cast approach. A detailed description of the LANL methodfor preparation of stable, reproducible mixed-potential sensors hasbeen published elsewhere [29,40]. Briefly, the YSZ electrolyte pow-der was dried in air at 100–150 ◦C for approximately 1 h. The dried

Actua

p((bmacwAiwipetcd

mawtelTwlbs

wsfltuhabaortetwfyhha

2

vtubOpiCmctt

Vk = Vk01

+ Ak[TOTAL]. (4)

J. Tsitron et al. / Sensors and

owder was mixed with solvents xylene, ethyl alcohol and fish oilBlown Menhaden) and ball-milled for 24 h. Plasticizers and bindersS-160 butyl benzyl phthalate, polyalkylene glycol and polyvinylutyral) were then added to the mixture, which was further ball-illed for 24 h. The mill was then discharged and de-aired for

pproximately 10 min at the reduced pressure of 20–25′′ of mer-ury. Prior to casting of the electrolyte slip, pieces of pure Au and Ptire (0.01′′ in diameter) were cut into 1′′ lengths and straightened.lternating pairs of Au and Pt were stuck to tape such that the spac-

ng between wire pairs was fixed; each set of Au/Pt electrode pairsould constitute a single device. This time-saving approach facil-

tated handling the electrodes during the upcoming time-criticalortions of sensor forming process, while maintaining the samelectrode separation and eventually the same electrode penetra-ion depth into the solid electrolyte. After de-gassing, the slurry wasast onto a Si-coated Mylar (G10JRM) carrier film using a standardoctor blade apparatus with a gap of approximately 0.05′′–0.2′′.

Once the tape was cast, it was allowed to partially dry for severalinutes. This resulted in a tape that was dry on the outside (facing

ir) but wet on the inside (in contact with the carrier film). This tapeas then turned upside down, exposing the wet side to air while

he dry side was in contact with the carrier film. The pre-fabricatedlectrode pairs were placed on top of this wet tape such that aength of 7.5 mm of each wire was embedded in the electrolyte.he tape was folded onto itself in order to enclose the electrodesithin the electrolyte. The remaining portions of the wires were

eft uncovered and exposed to air, and would eventually serve asoth electrode and signal-out leads attached to the data acquisitionystem.

Next, the electrolyte tape with the partially enclosed electrodesas allowed to air-dry fully, forming a mechanically stable, green

ensor body (the green sensor body refers to the unsintered andexible form of the tape before it is heated to a high tempera-ure to densify). Individual sensors were cut from the green tapesing a razor. The tape used to fix the electrode gap and to facilitateandling during the sensor-forming procedure was then removed,nd the sensor green bodies were placed into an alumina (Al2O3)oat and transferred to a tube furnace for removal of the organicsnd sintering using a multistep temperature profile. While previ-us sensor work of this nature conducted at LANL showed sensoresponse characteristics such as level of response and responseime to be dependent on sintering temperature [23], the use of goldlectrodes precludes sintering temperatures above 1064 ◦C. Thushe maximum sintering temperature used for sensors in this workas set at 1000 ◦C, above the lowest firing temperature (≈950 ◦C

or the Tosoh Corporation YSZ powders used in this work) that willield a device with sufficient mechanical stability to permit routineandling and flexing of wire electrodes. The YSZ porosity, whichas a strong effect on response times and sensor voltage [23], ispproximately 56%.

.2. Sensor response characterization

The sensor response or sensitivity (defined as the difference inoltage generated upon exposure to the test and base gases, respec-ively) was recorded using a Keithley 2400 source measurementnit with the Pt electrode connected to its positive terminal. Thease gas mixture was prepared by mixing pure streams of N2 and2 using MKS 1179A mass-flow controllers and the oxygen partialressure was maintained at 10% PO2 during single-gas sensor test-

ng (e.g. NH3 testing in the concentration range of 0–100 ppm). The3H6 interference (introduced to the test gas stream using another

ass-flow controller and 2500 ppm C3H6/balance N2 stock con-

entration) was mixed with the NH3/O2/N2 mixture at the inlet tohe furnace tube upstream of the sensor, imparting a small dilu-ion effect to O2 and NH3. A low-volume quartz furnace tube was

tors B 192 (2014) 283– 293 285

used that expands from 0.25 in. outside diameter (OD) at the coldends to approximately 0.75 in. OD for the portion surrounding thesensor and the platform supporting the sensor. The volume of themeasurement chamber is approximately 12 cm3.

2.3. Models of sensor voltage

Linear model. The response voltage of electrochemical sensorshas previously been modeled for a single compound, under somesimplifying assumptions, as a linear function of analyte gas con-centration [16]. Starting with this simple model, we assume thatmixtures of ammonia and propylene produce a response voltage inthe device which is a linear sum of the component gas concentra-tions, thus disregarding any interaction between analyte gases. Weexpress the sensor response voltage at each bias current setting kas:

Vk = Vk0 + Ak[NH3] + Bk[C3H6], (1)

or, equivalently,

Vk = Vk0 + (Ak + Bk˛)

([TOTAL]

1 + ˛

), (2)

where = [C3H6]/[NH3], [TOTAL] = [NH3] + [C3H6], and Ak, Bk are thelinear coefficients to be determined. Note that the advantage ofexpressing voltage in terms of and [TOTAL]i (i = 1, . . ., N˛, whereN˛ is the number of different total concentrations for a given ˛)rather than individual gas concentrations [NH3]i and [C3H6]i isthat the total concentration can be controlled experimentally, e.g.through a series of dilutions.1

Sensor response voltage readings were taken at 4 bias currentsettings (0, −1.5, −3.5, and −6 �A), with the sensor exposed todifferent mixtures of NH3 and C3H6. At each setting, a series ofmeasurements corresponding to different total concentrations wasextracted for several values of ˛:

[TOTAL]i (ppm) =

⎧⎪⎪⎪⎪⎨⎪⎪⎪⎪⎩

{0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100} for = 0,

{0, 40, 60, 120, 160, 180, 200} for = 1,

{0, 50, 100, 150} for = 1.5,

{0, 30, 90, 120, 150} for = 2,

{0, 40, 80, 120} for = 3,

{0, 20, 30, 60, 80, 90, 100} for ˛′ = 0

Note that in order to reconstruct [TOTAL]i for each series, only[TOTAL]ref needs to be predicted since

[TOTAL]i = [TOTAL]ref −i−1∑j=1

�cj, (3)

and �cj, the differences in total concentration between con-secutive measurements, are assumed to be known. We choose[TOTAL]ref = 200 ppm to be the total concentration at an arbitraryreference point. Once predictions for [TOTAL]ref and are made,they can be used to estimate [NH3]i and [C3H6]i for each individualmixture.

We use voltage measurements collected during sensor expo-sure to each test gas separately ( = 0 and ˛′ = 0) to estimate linearcoefficients Ak and Bk for each bias current k, as well as the offsetparameter Vk

0 = Vk01

+ Vk02

. When = 0, Eq. (2) becomes

1 Equivalently, we could have chosen to define the mixing ratio as˛′ = [NH3]/[C3H6]. This arbitrary choice of representation of the mixing ratio doesnot affect final predictions.

2 Actua

W

V

b

V

Eaasszs

bg

V

wttcsttviia

biaCbs

2

dl

L

wdanianicd

i

86 J. Tsitron et al. / Sensors and

hen ˛′ = 0, Eq. (2), written in terms of ˛′:

k = Vk0 + (Ak˛′ + Bk)

([TOTAL]1 + ˛′

),

ecomes

k = Vk02

+ Bk[TOTAL]. (5)

qs. (4) and (5) are used as input to RANSA (our algorithm describedt the end of Section 2), which predicts the average values of Ak, Bk,nd Vk

0 to be used in Eq. (2) (Table S1; least-squares fits yield veryimilar results). Note that although the baseline values have beenubtracted from all voltages, Vk

01and Vk

02are allowed to adopt non-

ero values as dictated by the global fits of non-linear curves totraight lines. In practice, these offsets are small (Table S1).

Nonlinear model. We can potentially improve our predictionsy incorporating interference between NH3 and C3H6 via a moreeneral nonlinear model:

k = Ck(˛)(

[TOTAL]1 + ˛

)pk(˛)

, (6)

here Ck(˛) and pk(˛) are unknown functions of ˛, with parame-ers to be estimated from the data. Since full Bayesian treatment ofhese functions is too laborious, we employ least-squares fits withross-validation. Specifically, we randomize each set of 140 mea-urements with the same [NH3]i and [C3H6]i and partition it into 5raining sets with 120 measurements each and 5 non-overlappingest sets with 20 measurements each. For each training set and eachalue of ˛, we find average Ck(˛) and pk(˛) using RANSA. Manualnspection of these results shows that, for each bias current k, Cks an approximately quadratic function of (Fig. S1), and pk is anpproximately linear function of between = 1 and = 3 (Fig. S2):

Ck(˛) = a′k

+ b′k

+ c′k˛2,

pk(˛) = ak + bk˛.(7)

We use a least-squares fit to estimate parameters ak, bk and a′k,

′k, c′

kon a given training set and test our model on the correspond-

ng test set by predicting and [TOTAL]ref for all available mixtures,gain using RANSA. Figs. S1 and S2 show variation in the predictedk(˛) and pk(˛), as well as the respective linear and quadratic fitsased on 5 training sets. For a complete list of model parameters,ee Tables S2–S4.

.4. Bayesian estimation of model parameters

We assume that the sensor voltage measurements are Gaussian-istributed around values predicted by the model, and compute the

og-likelihood of the data:

k = log P({Vk}|{�k}, ˛, [TOTAL]ref, �k)

= − 1

2 �2k

N∑l=1

(Vlk − V l

k)2 − N

2log(2� �2

k ), (8)

here k is the bias current index, Vlk({�k}, ˛, [TOTAL]ref, �) is pre-

icted voltage which depends on a set of parameters {�k}, {Vk} is set of voltage measurements V l

k(l = 1, . . ., N, where N is the total

umber of measurements with a given and bias current), and �k

s the noise parameter. The parameter set {�k} consists of Ak, Bk,nd Vk

0 for the linear model (Eq. (2)), and ak, bk, a′k, b′

k, c′

kfor the

onlinear model (Eqs. (6) and (7)). Note that predicted voltage Vlk

s a function of the measurement counter l because the total con-

entration, which depends on [TOTAL]ref through Eq. (3), may beifferent for each measurement.

The total log-likelihood is given by L =∑Ns

k=1Lk, where Ns = 4s the total number of bias current settings. In principle, the

tors B 192 (2014) 283– 293

log-likelihood can be used to estimate the posterior probability forall unknowns according to the Bayes rule [34]:

P({�}, ˛, [TOTAL]ref, { �}|{V })

= P({V }|{�}, ˛, [TOTAL]ref, { �})P({�})P(˛)P([TOTAL]ref)P({ �})P({V }) ,

(9)

where on the right-hand side the likelihood from Eq. (8) is multi-plied by the product of priors for each model parameter and dividedby evidence. {V } combines data from all bias current settings, { �}is a set of Ns noise parameters, and {�} is the union of all param-eter sets {�k}. P({�}) and P({ �}) are products of prior probabilitiesfor each parameter in the set. However, in practice we find it morereliable to estimate model parameters first in a separate calibrationstep:

P({�k}, �k|{Vk}, ˛, [TOTAL]ref)

= P({Vk}|{�k}, ˛, [TOTAL]ref, �k)P({�k})P( �k)

P({Vk}) . (10)

Specifically, with the linear model Eq. (4) is used to predict Ak andVk

01and, similarly, Eq. (5) is used to predict Bk and Vk

02. With the

nonlinear model, Eq. (6) is employed to estimate Ck(˛) and pk(˛).Note that and [TOTAL]ref are assumed to be known for the trainingdatasets. Finally, we compute ensemble averages � for all modelparameters � , and use them to predict and [TOTAL]ref:

P(˛, [TOTAL]ref, { �}|{V }, { �})

= P({V }|{ �}, ˛, [TOTAL]ref, { �})P(˛)P([TOTAL]ref)P({ �})P({V }) . (11)

We use uniform priors (invariant with respect to translations,x → x + a):

P(x) ={

1/(xmax − xmin) if x ∈ [xmin, xmax],

0 otherwise(12)

for {�} and [TOTAL]ref, and Jeffrey’s priors (invariant with respectto rescaling, x → ax) for and { �}:

P(x) ={

1/(x log(xmax/xmin)) if x ∈ [xmin, xmax],

0 otherwise(13)

In each case, xmin and xmax are chosen to encompass all plausiblesolutions and thus have no effect on final results.

We estimate all posterior probabilities by nested sampling [34]– a Bayesian Monte-Carlo (MC) technique that yields an ensem-ble of models from which the average value of each parameterand its standard deviation can be computed. Unlike other meth-ods such as MC sampling of the product of likelihood and priors,nested sampling allows us to keep track of the evidence, yieldingabsolute values of the posterior probability. Software used in thisstudy, called RANSA (Receptor Array Nested Sampling Algorithm),is available at http://olfaction.rutgers.edu.

3. Results and discussion

3.1. Effects of bias current application

Fig. 1 shows a schematic of the Au/YSZ/Pt tape-cast, mixed-potential sensor used in this work, along with a photograph of the

actual device. The sensor was placed into a tube furnace and testedat a series of temperatures in order to determine optimum per-formance specified by the tradeoff of NH3 sensitivity and responsetime. Fig. 2(a) shows the open-circuit response of the sensor for

J. Tsitron et al. / Sensors and Actuators B 192 (2014) 283– 293 287

Fig. 2. Au/YSZ/Pt mixed-potential sensor response to NH3 is subject to HC interference. (a) Open-circuit (0 �A) response of the sensor to 100 ppm of NO2, NO, NH3, and C3H6.(b) Same as (a), at −3 �A current bias. (c) Sensor response to 0, 25, 50, 75, and 100 ppm of NH3 and equivalent amounts of NO2 and NO at −3 �A applied current bias. (d)Sensor response to 100 ppm of NO2, NO, NH3, and C3H6 at various levels of positive and negative current bias. Note that the baseline voltage (measured with no test gaspresent) is subtracted from the voltage measured with 100 ppm of each test gas. All measurements were taken with sensors at 575 ◦C and in 10% O2/balance N2, flowing at5

iWtes1bt

tfettbadhtiid

00 ml/min.

ndividual 5-min exposures to 100 ppm of NO, NO2, NH3, and C3H6.hile the sensor has relatively little sensitivity to NO and NO2,

here is substantial cross-sensitivity to C3H6. Fig. 2(b) shows theffect of the application of a −3 �A bias current on the relativeelectivity for this device. We observe that while the response to00 ppm NO2 has been removed, the response to 100 ppm NO haseen augmented and that to C3H6 reduced by roughly 30% relativeo NH3.

Fig. 2(c) shows that in the absence of hydrocarbon gases,his sensor could serve as a highly selective ammonia sensoror SRC applications, as it exhibits limited sensitivity to NO andspecially NO2. Fig. 2(d) shows voltage response to 100 ppm ofhe four test gases as the bias current is varied from −4.5 �Ao +3.5 �A. With the application of increasing negative currentias, the NO, NO2, and C3H6 levels decrease with respect to themmonia response; however, the voltage for 100 ppm of C3H6oes not decrease below −4.5 �A and therefore this particularydrocarbon remains a substantial source of interference with

he ammonia signal. Thus the ability to alter response selectiv-ty with the Au/YSZ/Pt mixed-potential sensor is limited by HCnterference. This fact, taken together with preliminary sensoresign studies of the effect of electrode geometry and the ratio of

bulk to interfacial resistance on Au/YSZ/Pt sensor response [33],shows that some level of undesirable cross-interference with non-target gas species such as C3H6 will be inevitable with this sensorconstruct.

3.2. Deconvoluting ammonia response in the presence ofhydrocarbon interference

In light of the unavoidable HC interference discussed above, wecarried out a series of experiments aimed at collecting NH3/C3H6interference data for subsequent mathematical modeling. Eachseries of measurements was performed at four different bias cur-rent settings. Since in this regime sensor selectivity and sensitivitydepend on the magnitude of the bias current (Fig. 2(d)), using foursettings is equivalent to creating an array of four sensors, each ofwhich produces a different response to a given mixture of NH3 andC3H6. Fig. 3(a) shows the baseline-corrected sensor response to var-ious levels of NH3 in the presence of 0, 20, 30, 60, 80, 90 and 100 ppm

of the interfering gas, C3H6, with an applied bias current of −6 �A(data for bias current settings of 0, −1.5, and −3.5 �A is shownin Figs. S3(a), S4(a), and S5(a), respectively). These measurementsare used as input to a Bayesian inference algorithm called RANSA

288 J. Tsitron et al. / Sensors and Actuators B 192 (2014) 283– 293

Fig. 3. Sensor response to mixtures of NH3 and C3H6. (a) Sensor response voltage plotted as a function of time and [C3H6]. [NH3] levels increase from 0 to 100 ppm in 10 ppmincrements in the same time course, and each curve corresponds to a different level of [C3H6], as indicated in the legend. (b) Sensor response vs. total concentration [TOTAL],sorted by the mixing ratio (squares). Lines with open circles represent linear model predictions (Eq. (2)). Raw interference data in (a) corresponds to = 0, 1, 1.5, 2, 3, and˛′ ≡ [NH3]/[C3H6] = 0. Each data point is an average of the approximately 140 voltage readings for each pair of concentrations in (a) (standard errors are omitted for clarity).For example, the 200 ppm data point for = 1 (blue square), is an average of all measurements taken when the sensor was exposed to 100 ppm of NH3 and 100 ppm of C3H6,i.e., the black line in (a) from t = 51 min to t = 54 min. Note that = 0 and ˛′ = 0 series of measurements correspond to individual gases rather than mixtures. (c) Same data asin (b) (squares), with the ˛′ = 0 case omitted. Lines with open circles represent non-linear model predictions (Eq. (6)). With the exception of the = 0 (black) curve, Ck andpk (cf. Eq. (6)) are functions of ¯ , the mean value of as predicted across five training/test data subsets (see Section 2). The curve coefficients in Eq. (7) (ak, bk; a′

k, b′

k, c′

k) are

also averaged over five training/test subsets. In the case of = 0, Ck is estimated as before but pk( = 0) is the mean value inferred by RANSA in the calibration step ratherthan predicted by a least-squares fit (see Fig. S2 and Table S2). In (b) and (c), each predicted curve is color-coded to correspond to different mixing ratios, as defined in thelegends. Note that [TOTAL] = 0 measurements and predictions were omitted for clarity, except for the = 0 and ˛′ = 0 cases in (b) which were used to calibrate the linearmodel. For all panels, the bias current was set at −6 �A, the baseline voltage (voltage measured with no test gases present) was subtracted from all voltage measurements,and all measurements were taken at 575 ◦C and in 10% O2/balance N2, flowing at 500 ml/min. (For interpretation of the references to color in this figure legend, the reader isr

(N

dwccSv

eferred to the web version of the article.)

Receptor Array Nested Sampling Algorithm) [39], which predictsH3 and C3H6 concentrations in each mixture.

Linear model. Since the magnitude of hydrocarbon interferenceepends on the relative concentration of C3H6 with respect to NH3,e find it convenient to replot the data in terms of the relative

oncentration (or mixing ratio) ≡ [C3H6]/[NH3], and the total con-entration [TOTAL] ≡ [NH3] + [C3H6] (Figs. 3(b), S3(b), S4(b), and5(b)). Each series of points represents baseline-corrected sensoroltage measurements at a given bias current, as a function of the

total mixture concentration [TOTAL], for a given value of (weonly consider gas mixtures for which four or more total gas con-centrations are available, see Section 2). We see that, to a firstapproximation, the response of each sensor to a given mixture islinear.

The linear model assumes that the total response to a mixtureof gases is simply a sum of responses to each individual gas in themixture (Eq. (1)). Thus, to fit mixture data to a linear model, we firstdetermine its parameters (Vk

0 , Ak, and Bk, where k = 1, . . ., 4 labels

Actuators B 192 (2014) 283– 293 289

btSpaeacdeda

imtbtlcia0

wdaotmeawe

fuTtpcSnsttCbme

tuwpes[lmbntbim

0

1

1.5

2

3

α

predictedactual

α = 0 α = 1 α = 2 α = 3 α = 1.5

0

50

100

150

200

predictedactual

[TOTAL]ref

α = 0 α = 1 α = 2 α = 3 α = 1.5

a

b

Fig. 4. Prediction of gas concentrations in mixtures of NH3 and C3H6. Using a non-linear model (Eq. (6)) and a Bayesian nested sampling algorithm RANSA [39], wepredict means and standard deviations of the mixing ratio, (a), and the total mix-ture concentration at the 200 ppm reference point, [TOTAL]ref (b) for five different

J. Tsitron et al. / Sensors and

ias current settings; cf. Eqs. (4) and (5)) by using sensor responseso each separate gas (black and red squares in Figs. 3(b), S3(b), S4(b),5(b)). Although we obtain full Bayesian estimates for all modelarameters by applying RANSA to Eq. (10), we only employ theirverage values in subsequent predictions, as integrating over thensemble of models would be prohibitively expensive. We thenpply RANSA to Eq. (11) to predict the mixing ratio and the totaloncentration at the reference point [TOTAL]ref (equivalent to pre-icting absolute concentrations of NH3 and C3H6, see Section 2) forach of the four mixtures. Despite neglecting all nonlinearity in theata, we nonetheless do reasonably well at predicting individualnalyte concentrations (Fig. S6).

Nonlinear model. Although the sensor response is approx-mately linear, interference between NH3 and C3H6 becomes

ore prominent at higher gas concentrations, making the addi-ivity approximation less accurate. Interestingly, cross-interferenceetween the two gases depends on both the mixing ratio andhe bias current. For example, at 0 �A the voltage response is sub-inear, indicating a preferential response to one of the two gasomponents (Fig. S3(c)). This trend is gradually reversed with thencrease in the negative current bias, resulting in voltage responsest −6 �A that are inverted compared with their counterparts at

�A (Figs. S4(c), S5(c) and 3(c)).To model nonlinear sensor responses, we employ a straightfor-

ard generalization of Eq. (2) shown in Eq. (6). The nonlinear modelepends on two independent parameters, Ck(˛) and pk(˛), whichre functions of both the mixing ratio and the bias current, asbserved in the data. To predict these functions and thus modelhe sensor response over a range of ˛, we first divide all measure-

ents into 5 training/test subsets as described in Section 2. Forach training subset, we predict Ck(˛) and pk(˛) using RANSA, forll available values of ˛: 1.0, 1.5, 2.0, and 3.0. In this calibration step,e assume and [TOTAL]ref to be known, and employ Eq. (10) to

stimate means and standard deviations of the model parameters.Next, we employ polynomial fits to infer Ck and pk as explicit

unctions of (Eq. (7); cf. Figs. S1 and S2) for each training subset,sing RANSA-predicted mean values of Ck(˛) and pk(˛) as input.hese steps yield a set of 5 models for Ck(˛) and pk(˛), allowing uso cross-validate our predictions. Optionally, a single model can beroduced by averaging the polynomial parameters of Ck(˛) (a′

k, b′

k,

′k) and pk(˛) (ak, bk) over 5 subsets of training data, as in Fig. 3(c).

ince the polynomial fits are done by least squares, they represent aon-Bayesian step which was implemented because we do not haveufficient data for full Bayesian inference (Eq. (9)). This necessitateshe use of cross-validation. We note that, unlike the linear model,he nonlinear model is limited to a certain range of ˛’s becausek(˛) and pk(˛) predictions cannot be reliably extrapolated too fareyond the set of ˛’s used in the calibration step. Thus the nonlinearodel should be trained using a set of mixing ratios that cover the

ntire range of interest in subsequent applications.Models trained on 5 training data subsets are then tested on their

est set counterparts. RANSA is utilized again to predict individ-al analyte concentrations (or, more specifically, and [TOTAL]refhich encode the same information, see Section 2) using Ck(˛) and

k(˛) functions as input to Eq. (11) (Fig. 4). Note that, as in the lin-ar model, these predictions combine data from all bias currentettings. We see a sizable improvement in our predictions of andTOTAL]ref when the nonlinear model is employed: the mean abso-ute error in predicting for four mixtures is 0.80 with the linear

odel, but is reduced to 0.29 with the nonlinear model. The errorecomes 0.23 when the = 0 case is included in the average foronlinear predictions. Note that this case cannot be used to test

he linear model because = 0 (single gas) data was used to cali-rate it. Similarly, the mean absolute error in predicting [TOTAL]ref

s 2.6 ppm with the linear model, and 2.1 ppm with the nonlinearodel when the = 0 case is included. The essential differences

values of from Fig. 3(c). All means and standard deviations are averaged over fivetraining/test data subsets (see Section 2).

between linear and nonlinear models also manifest themselves inthe quality of predictions shown in panels (b) and (c) of Figs. 3and S3–S5. In the nonlinear case, the model predictions employpredicted mixing ratios and predicted Ck and pk functions, evenin the = 0 case. In contrast, = 0 and ˛′ = 0 data were used tocalibrate the linear model as described in Section 2, and thus lin-ear model predictions closely reproduce the data in those twocases.

3.3. Minimum data requirements for robust concentrationpredictions

Our computational framework for predicting gas analyte con-centrations consists of two steps: the calibration step, in which theparameters of the model are inferred in a laboratory setting using anarray of sensors, and the prediction step, in which the same sensorsetup is employed to predict and [TOTAL]ref (and thus absoluteconcentrations of each constituent analyte) in real-world condi-

tions. (In this study, all measurements were done in the lab, withsome of the data set aside to mimic real-world measurements.)While we can assume that sufficient training data is available forgas mixtures of interest for subsequent practical applications, the

290 J. Tsitron et al. / Sensors and Actuators B 192 (2014) 283– 293

Fig. 5. Robustness of gas concentration predictions with respect to the number of measurements in training and test datasets. (a) Average absolute error �(˛) for predictionsof ˛, plotted as a function of the number of points with different total concentrations. ‘Range 1’ and ‘Range 2’ are defined in the text. Mean values of estimated by RANSAwere averaged across 5 training datasets. The average absolute error is given by the absolute magnitude of the difference between this average and the exact value = 1.(b) Same as (a) but the average absolute errors are shown for predictions of [TOTAL]ref. The exact value is [TOTAL]ref = 200 ppm. (c) Means and standard deviations of ˛predicted by RANSA after omitting the mixture for which the predictions were made from the training data, and recalibrating the model. All means and standard deviationsare averaged over 5 training subsets. (d) Same as (c) but for means and standard deviations of [TOTAL]ref. Predictions in (a)–(d) use data from all 4 bias current settings. (Fori b vers

nid

atiuotpdmmsipfvwlbt

nterpretation of the references to color in the text, the reader is referred to the we

umber of measurements taken in the prediction step must be min-mized in order to improve reaction times and efficiency of theevice.

Our predictions in Fig. 4 were based on 20 independent volt-ge readings for each mixture and each bias current setting (forhe nonlinear model which was previously calibrated on 5 train-ng datasets as described above). Since in practice a single devicender different bias current settings rather than a physical arrayf sensors is likely to be employed, it is important to minimizehe number of measurements in each series of readings. Sur-risingly, we find that the quality of our predictions does notecline significantly when we gradually reduce the number ofeasurements from 20 to 1 for each k, ˛, and [TOTAL]. Thisay be due to the fact that each measurement, taken at the

teady-state flow of the gas mixture, conveys roughly the samenformation to the algorithm. Indeed, the mean absolute error inredicting all ˛’s is 0.38 for 2 measurements, 0.22 for 5, and 0.28or 20. The mean absolute error is computed by first averaging meanalues of predicted by RANSA for each mixture using 5 test sets

ith the number of points indicated above as input. Next, the abso-

ute error is computed as the absolute magnitude of the differenceetween predicted and exact values, and averaged over all 5 mix-ures (with = 0.0, 1.0, 1.5, 2.0, 3.0). Similarly, the mean absolute

ion of the article.)

error in predicting [TOTAL]ref is 2.54 for 2 measurements, 2.16 for 5,and 2.36 for 20. Note that, as expected, standard deviation in eachindividual RANSA prediction of and [TOTAL]ref increases as thenumber of measurements is reduced, although the uncertaintiesremain small.

Our algorithm depends on voltage measurements taken at morethan one value of the total concentration for a given mixture (cf. Eq.(3)). In real-world applications, controlled dilution of the mixtureof ammonia and hydrocarbons in a neutral buffer such as fresh airwill make the device more complex to build and operate. Thus, weinvestigate how the quality of predictions depends on the num-ber of different total concentrations or, equivalently, the numberof dilutions for the same mixture. Since among 4 mixtures with

/= 0 we have the most data for the = 1 case (seven total concen-trations; blue squares in Figs. 3(c), S3(c), S4(c), and S5(c)), we focuson this dataset for our analysis. We start with two points, corre-sponding to [TOTAL] = 0 ppm and [TOTAL] = 40 ppm (‘Range 1’), andthen gradually extend the range by including points with higherconcentrations until the full dataset is recovered. We predict and

[TOTAL]ref using the nonlinear model, and plot absolute errors ofour predictions as blue crosses in Fig. 5(a) and (b). In each individ-ual calculation, we use the nonlinear model trained on 5 completecalibration datasets.

J. Tsitron et al. / Sensors and Actua

Fig. 6. Prediction accuracy as a function of the number of bias current settings. (a)Each point represents the average predicted value of ˛. First, mean ˛’s are obtainedfor each of the 5 training subsets and for each combination of sensors. These mean˛’s are averaged over the training subsets and, next, over the sensor combinations.The error bars represent standard deviations with respect to different sensor com-binations. Exact values are shown as horizontal dashed lines of the same color. Lightgray bars indicate the number of bias current settings (predictions are slightly offsetfor clarity). (b) Same as (a) but for [TOTAL]ref. (For interpretation of the referencesto color in this figure legend, the reader is referred to the web version of the article.)

t[sistccf[mncs

absolute concentration of each gas. Our Bayesian approach allowsus to build an ensemble of models that explain the data (rather than

We repeat the calculation starting at the high concentra-ion end (‘Range 2’), first using two points corresponding toTOTAL] = 200 ppm and [TOTAL] = 180 ppm and then adding mea-urements with lower concentrations until all seven points arencluded. Errors in and [TOTAL]ref for this set of calculations arehown as red circles in Fig. 5(a) and (b) respectively. We observehat the quality of predictions strongly depends on which totaloncentrations are included in the analysis. It appears that low-oncentration measurements (Range 1), where the contributionsrom both gases are approximately additive, can be used to inferTOTAL]ref more accurately than ˛. In contrast, high-concentration

easurements, where cross-interference effects are more promi-ent, appear to be more beneficial for the inference of ˛. In both

ases, average errors exhibit non-uniform trends but ultimatelytart to converge as the range is extended. Thus the number of

tors B 192 (2014) 283– 293 291

times each mixture is diluted should be chosen carefully dependingon the level of accuracy desired for each predicted variable.

Next, we check the ability of our algorithm to predict and[TOTAL]ref for mixtures that were not part of the training dataset.This ability is crucial because our framework must yield reasonablyaccurate predictions for a range of ˛’s rather than just the discreteset of mixtures on which it was trained. To this end, we leave out alldata corresponding to one of the /= 0 ( = 1, 1.5, 2, or 3), and refitthe nonlinear model using the remaining 3 mixtures, plus the datafor = 0. The model is then used to predict the total gas concentra-tions in the mixture that was left out of the calibration step (Fig. 5(c)and (d)). We see that, encouragingly, the errors for = 1.5 are rea-sonably small, likely because the data for both = 1 and = 2 wereavailable to train the model. On the other hand, our ability to extrap-olate the range of is limited, as manifested by the = 3 example.Thus the model calibration step should be based on training datathat covers the entire range of ˛’s to be expected in subsequentapplications, with reasonably fine spacing between adjacent valuesof ˛.

Finally, we evaluate the benefit of combining data collected atmultiple bias current settings. We monitor errors in our predic-tions as additional sensors are incorporated into the array. Usingfewer current settings would simplify construction and operationof the device. Here, we use the nonlinear model calibrated earlieron all 5 training datasets and all 4 bias current settings. However,instead of combining the data from all 4 bias currents, we estimatethe errors in and [TOTAL]ref as the number of bias current settingsis increased from 1 to 4. With 4 settings, there are 4 combinationswith a single sensor in the array, 6 combinations with two sensors, 4with three, and a single one with the full dataset. For each combina-tion, we predict and [TOTAL]ref (Fig. 6). We see that, as expected,the accuracy of our inference improves as additional bias currentsettings are used to collect the measurements. However, if neces-sary, reasonable accuracy can be achieved even with 2–3 sensors,especially if the total concentration is of primary interest. Over-all, the differential sensitivity and selectivity of the device whichresults from the use of multiple bias currents can be exploited forimproved discrimination of gas mixtures.

4. Conclusions

Modern diesel engines require continuous monitoring of NOx

and NH3 levels in tailpipe emissions to ensure regulatory com-pliance and minimize air pollution. Au/YSZ/Pt mixed-potentialsensors represent a promising technology which could address thisneed. The sensors are compact and rugged (Fig. 1) and, moreover,their sensitivity and selectivity toward target gases can be tuned byapplying a bias current (Fig. 2). Thus a single sensor under differentbias current settings can be used to create an entire array of sensors,yielding more robust and accurate predictions of gas concentra-tions in mixtures. Unfortunately, currently available zirconia-basedelectrochemical sensors exhibit cross-interference with hydrocar-bons such as propylene (C3H6) (Fig. 2(d)). Thus sensor voltagereadings are not simply proportional to the amount of target gasin diesel exhaust.

Here we show that target gas concentrations can nonetheless beinferred from sensor readouts if data is collected under several biascurrent settings. We focus on mixtures of NH3 (target gas) and C3H6(interfering gas), and develop a Bayesian framework for predicting˛, the ratio of concentrations between the two gases, and the sum ofthe two concentrations. This information is equivalent to predicting

a single maximum-likelihood model), and thus evaluate both themean value of each parameter of interest and its uncertainty. The

2 Actua

mcsatm

itWftatntS

as(ftsdtclatiteegleac

gconasNgeioibtfove

tdocstb

[

[

[

[

[

[

[

[

[

[

[

[

92 J. Tsitron et al. / Sensors and

ethod is not limited to Au/YSZ/Pt mixed-potential sensors andan be employed to infer concentrations for more than two gasesimultaneously. We expect our approach to be widely applicable inutomotive as well as other industrial applications which rely onhe accurate knowledge of gas concentrations in complex chemical

ixtures.The simplest model we have investigated assumes that the gases

n the mixture do not interfere with one another, and hence theotal voltage is a sum of contributions from each individual gas.

e train the parameters of this linear model using sensor readingsrom individual gases, and then apply it to predict gas concentra-ions in mixtures. The linear model is easy to set up and calibrate,nd has few fitting parameters. Its accuracy does not depend onhe relative proportions of the gases in the mixture. However, iteglects cross-interference between gases at higher gas concen-rations, which causes nonlinearities in sensor response (Figs. 3(c),3(c), S4(c), and S5(c)).

In order to check whether modeling this effect improves theccuracy of our predictions, we have built a nonlinear model as atraightforward generalization of the linear model (cf. Eqs. (2) and6)). The nonlinear model has fitting parameters that are explicitunctions of (see Section 2), and therefore must be trained on datahat uniformly covers the range of mixtures to be encountered inubsequent applications. This makes the training procedure moreata-intensive, and limits the applicability of the model. In addi-ion, the model is not fully Bayesian and as a result needs to beross-validated. The degree of polynomials used in curve fittingikely depends on the application (sensor types, range of ˛, etc.)nd needs to be re-examined in each case. Despite these difficul-ies, prediction accuracy improves when nonlinearities are takennto account, which justifies using the nonlinear model in situa-ions where it can be first calibrated in a laboratory setting usingxtensive training datasets, and then deployed in the field. Inter-stingly, the nonlinear model captures cross-interference betweenases without explicit knowledge of reaction activity at molecularevel. Note that in this proof-of-concept study, we used the samexperimental setup (sensor position, flow rate, gas source, temper-ture, etc.) to both train the models and make predictions of gasoncentrations in mixtures.

Using our sensors and models in real-world conditions will bereatly facilitated by making as few measurements as possible. Toheck what the minimal requirements are for the deployment ofur system, we have carried out four tests in which we reduce theumber of test datapoints in various ways and monitor predictionccuracy of the nonlinear model. First, we kept all 4 bias currentettings and the entire available range of total concentrations ofH3 and C3H6, but reduced the number of datapoints for eachiven total concentration and bias current setting. We found that,ncouragingly, just 1–2 datapoints are necessary, which shouldmprove response times. More crucially, our algorithm dependsn the availability of data for several total concentrations, whichn practice entails diluting the mixture of interest in a neutraluffer. Since this step adds complexity to the system, we checkedhe minimum number of total concentrations that are necessaryor acceptably accurate predictions. It appears that the numberf dilutions can be reduced by judicious choice of neutral bufferolumes, especially if only or the total concentration are of inter-st.

Next, we tested the robustness of our framework with respecto extrapolating to the values of which do not appear in trainingata. We found that prediction accuracy is sensitive to such extrap-lations, which argues for carefully choosing the range in the

alibration step. Finally, we checked whether the number of sen-ors (i.e., bias current settings) can be reduced below four. We sawhat although adding sensors to the array is clearly beneficial, it maye acceptable to use 2–3 or even a single sensor without too much

[

[

tors B 192 (2014) 283– 293

accuracy loss. Overall, we conclude that it is possible to achieveacceptable accuracy levels, system complexity, and reaction timeswith our sensors and algorithms.

Acknowledgements

This research was funded by the US DOE, EERE, Vehicle Technol-ogy Programs. The authors wish to thank Technology DevelopmentManager Roland Gravel. A.V.M. acknowledges support from anAlfred P. Sloan Research Fellowship.

Appendix A. Supplementary Data

Supplementary data associated with this article can be found, inthe online version, at http://dx.doi.org/10.1016/j.snb.2013.10.115.

References

[1] M. Kampa, E. Castanas, Human health effects of air pollution, Environ. Pollut.151 (2008) 362–367.

[2] P. Scheepers, R. Bos, Combustion of diesel fuel from a toxicological perspective,Int. Arch. Occup. Environ. Health 64 (1992) 149–161.

[3] G.S. Sandhu, H.C. Frey, Real-world measurement and evaluation of duty cycles,fuels, and emissions control technologies of heavy-duty trucks, Transport Res.Rec. 2270 (2012) 180–187.

[4] R. Burch, J. Breen, F. Meunier, A review of the reduction of NOx with hydrocar-bons under lean-burn conditions with non-zeolite oxide and Pt group metalcatalysts, Appl. Catal. B 39 (2002) 283–303.

[5] S. Noroozi, Urea enhances safety in SCR applications, Power Eng. 97 (1993)28–29.

[6] G.H. Abd-Alla, Using exhaust gas recirculation in internal combustion engines:a review, Energy Convers. Manage. 43 (2002) 1027–1042.

[7] A. Maiboom, X. Tauzia, J.-F. Hetet, Experimental study of various effects ofexhaust gas recirculation (EGR) on combustion and emissions of an automotivedirect injection diesel engine, Energy 33 (2008) 22–34.

[8] R. Anand, N.V. Mahalaskmi, Simultaneous reduction of oxides of nitrogen andsmoke emissions by using EGR combined with particulate trap in a direct injec-tion diesel engine, in: Proc. 2006 Spring Tech. Conf. ASME Internal CombustionEngine Division, 2006, pp. 209–216.

[9] F. Schafer, R. van Basshuysen, Reduced Emissions and Fuel Consumption inAutomobile Engines, Springer-Verlag, New York, 1995.

10] P. Sekhar, E. Brosha, R. Mukundan, W. Li, M. Nelson, P. Palanisamy, F.H. Gar-zon, Application of commercial automotive sensor manufacturing methodsfor NOx/NH3 mixed potential sensors for on-board emissions control, Sens.Actuators B: Chem. 144 (2010) 112–119.

11] P. Forzatti, Present status and perspectives in de-NOx SCR catalysis, Appl. Catal.222 (2001) 221–236.

12] M. Hsieh, J. Wang, Development and experimental studies of a control-orientedSCR model for a two-catalyst urea-SCR system, Control Eng. Practice 19 (2011)409–422.

13] M. Hsieh, J. Wang, Sliding-mode observer for urea-selective catalytic reduction(SCR) mid-catalyst ammonia concentration estimation, Int. J. Automotive Tech.12 (2011) 321–329.

14] J. Visser, R. Soltis, Automotive exhaust gas sensing systems, IEEE Trans. Instrum.Measur. 50 (2001) 1543–1550.

15] J.W. Fergus, Sensing mechanism of non-equilibrium solid-electrolyte-based chemical sensors, J. Solid State Electrochem. 15 (2010)971–984.

16] F.H. Garzon, R. Mukundan, E. Brosha, Solid-state mixed potential gas sen-sors: theory, experiments and challenges, Solid State Ionics 136–137 (2000)633–638.

17] N. Miura, T. Koga, M. Nakatou, P. Elumalai, M. Hasei, Electrochemical NOx

sensors based on stabilized zirconia: comparison of sensing performances ofmixed-potential-type and impedancemetric NOx sensors, J. Electroceram. 17(2006) 979–986.

18] S. Zhuiykov, N. Miura, Development of zirconia based potentiometric NOx sen-sors for automotive and energy industries in the early 21st century: what arethe prospects for sensors, Sens. Actuators B: Chem. 121 (2007) 639–651.

19] B. Timmer, W. Olthuis, A. van den Berg, Ammonia sensors and their applications– a review, Sens. Actuators B: Chem. 107 (2005) 666–677.

20] X. Liang, T. Zhong, H. Guan, F. Liu, G. Lu, B. Quan, Ammonia sensors based onNASICON and Cr2O3 electrodes, Sens. Actuators B: Chem. 136 (2009) 479–483.

21] W.J. Fleming, Physical principles governing nonideal behavior of the zirconiaoxygen sensor, J. Electrochem. Soc. 124 (1977) 21–28.

22] P. Pasierb, M. Rekas, Solid-state potentiometric gas sensors-current status andfuture trends, J. Solid State Electrochem. 13 (2009) 3–25.

23] P.K. Sekhar, E.L. Brosha, R. Mukundan, M.A. Nelson, D. Toracco, F.H. Garzon,Effect of YSZ sintering temperature on mixed potential sensor performance,Solid State Ionics 181 (2010) 947–953.

Actua

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

B

JlBisacm

CMfrLe

Pp

University. In 2007 he joined the Department of Physics and Astronomy at RutgersUniversity, where he is now an Associate Professor. He is also a faculty member

J. Tsitron et al. / Sensors and

24] R. Mukundan, E.L. Brosha, F.H. Garzon, Mixed potential hydrocarbon sensorsbased on a YSZ electrolyte and oxide electrodes, J. Electrochem. Soc. 150 (2003)H279–H284.

25] R. Mukundan, E.L. Brosha, F.H. Garzon, Electrodes for solid state gas sensor.Patent, 08 2003. 6,605,202.

26] F.H. Garzon, E.L. Brosha, R. Mukundan, Method for forming a potential hydro-carbon sensor with low sensitivity to methane and CO. Patent, 12 2003.6,656,336.

27] E.L. Brosha, R. Mukundan, R. Lujan, F.H. Garzon, Mixed potential NOx sensorsusing thin film electrodes and electrolytes for stationary engine type applica-tions, Sens. Actuators B: Chem. 119 (2006) 398–408.

28] E.L. Brosha, R. Mukundan, D.R. Brown, Q.X. Jia, R. Lujan, F.H. Garzon, Techniquesfor the thin film growth of La1−xSrxCrO3 for solid state ionic devices, Solid StateIonics 166 (2004) 425–440.

29] R. Mukundan, F.H. Garzon, E.L. Brosha, Electrodes for solid state gas sensor.Patent, 05 2007. 7,214,333.

30] C. Boyer, S. Zemskova, P. Park, L. Balmer-Millar, D. Endicott, S. Faulkner, DOEEERE Report, 2003.

31] R. Mukundan, K. Teranishi, E.L. Brosha, F.H. Garzon, Nitrogen oxide sensorsbased on yttria-stabilized zirconia electrolyte and oxide electrodes, Elec-trochem. Solid-State Lett. 10 (2) (2007) J26–J29.

32] D.L. West, F.C. Montgomery, T.R. Armstrong, Total NOx sensing elements withcompositionally identical oxide electrodes, J. Electrochem. Soc. 153 (2006)H23–H28.

33] N. Miura, G. Lu, M. Ono, N. Yamazoe, Selective detection of NO by using anamperometric sensor based on stabilized zirconia and oxide electrode, SolidState Ionics 117 (1999) 283–290.

34] D. Sivia, J. Skilling, Data Analysis: A Bayesian Tutorial, Oxford University Press,Oxford, 2006.

35] S. Firestein, How the olfactory system makes sense of scents, Nature 413 (2001)211–218.

36] K. Touhara, L.B. Vosshall, Sensing odorants and pheromones with chemosen-sory receptors, Annu. Rev. Physiol. 71 (2009) 307–332.

37] K.J. Albert, N.S. Lewis, C.L. Schauer, G.A. Sotzing, S.E. Stitzel, T.P. Vaid, D.R. Walt,Cross-reactive chemical sensor arrays, Chem. Rev. 100 (2000) 2595–2626.

38] A.T. Wright, E.V. Anslyn, Differential receptor arrays and assays for solution-based molecular recognition, Chem. Soc. Rev. 35 (2006) 14–28.

39] J. Tsitron, A. Ault, J. Broach, A. Morozov, Decoding complex chemical mixtureswith a physical model of a sensor array, PLoS Comput. Biol. 7 (2011) e1002224.

40] F.H. Garzon, E.L. Brosha, R. Mukundan, Solid state ionic devices for combustiongas sensing, Solid State Ionics 175 (2004) 487–490.

iographies

ulia Tsitron earned a combined B.A./M.A. degree in Physics from Hunter Col-ege, City University of New York (2008), and is pursuing a Ph.D. in Computationaliology and Molecular Biophysics at Rutgers University. Her research interests

nclude developing mathematical models and software tools for analysis of sen-or response, robustness, and optimization, to be used across a broad range ofpplications. She is co-inventor (patent pending) of Chemosensory arrays, a yeastell-based sensor array and method of discriminating ligands in complex chemicalixtures.

ortney Kreller received her B.S. in Chemical Engineering from the University ofaryland, Baltimore County in 2004. She received her Ph.D. in Chemical Engineering

rom the University of Washington in 2011. Dr. Kreller is currently a post-doctoralesearcher in the Materials Physics and Applications division at Los Alamos National

aboratory. Her research interests include electrochemical sensors, fuel cells, andlectrosynthesis of fuels.

raveen Sekhar is an assistant professor in the School of Engineering and Com-uter Science at Washington State University Vancouver, WA since 2011. He was a

tors B 192 (2014) 283– 293 293

postdoctoral research associate in the Materials Physics and Applications division(MPA-11) at the Los Alamos National Laboratory for two years. He received his B.E.(distinction) from Coimbatore Institute of Technology, India, in 2001. He receivedhis Masters and Ph.D. in Electrical Engineering from the University of South Florida,Tampa, FL in 2005 and 2008, respectively. His dissertation focused on the syn-thesis, characterization and applications of silica nanowires won the ‘OutstandingDissertation’ award. Currently, his research interests include the developmentof electrochemical gas sensors and biosensors in addition to materials sci-ence and characterization, impedance spectroscopy, microfabrication, and appliedstatistics.

Rangachary Mukundan graduated from the University of Roorkee (India) with aBachelor’s degree in Metallurgical Engineering in 1991. He was a research fellow atthe University of Pennsylvania, Philadelphia, Pennsylvania, and received his Ph.D.in Materials Science and Engineering in February 1997. His thesis titled “Characteri-zation of Mixed-Conducting Barium Cerate-Based Perovskites for Potential Fuel CellApplications” was awarded the S.J. Stein Prize for superior achievement in the field ofnew or unique materials in electronics. Mukundan is a staff member in the MaterialsPhysics and Applications division (MPA-11) at the Los Alamos National Labora-tory. Currently, his research interests include fuel cell materials, electrochemicalgas sensors, proton-conductors, and permeation membranes.

Fernando Garzon is the Team Leader for materials chemistry in Materials Physicsand Applications division (MPA-11) at the Los Alamos National Laboratory. Dr.Garzon received his BSE in Metallurgy and Materials Science, and his Ph.D. inMaterials Science and Engineering from the University of Pennsylvania (Philadel-phia). His research interests include: the development of micro-electrochemicalsensors, polymer fuel cell technology, solid oxide fuel cell technology, the ther-mochemistry of electronically conducting transition metal oxides, thin film growthof oxide materials and ceramic membrane technology for sensor, gas separa-tion and fuel cell applications and the characterization of nano-scale materialsby advanced X-ray scattering methods. Dr. Garzon has co-authored over 100peer-reviewed scientific publications and made numerous invited conference pre-sentations. Research highlights include: the first experimental determination ofthe thermodynamic metastability of high temperature superconductors publishedin Science, the development of very low surface resistance superconductor thinfilms for microwave applications, the invention of non-porous ceramic hydrogenseparation membranes and the development of sulfur tolerant solid oxide fuelcells. He is also the co-inventor of an R&D 100 award-winning high tempera-ture, combustion control sensor currently being licensed to industry, and a newclass of solid-state gas sensors. He holds six patents in electrochemical technol-ogy and has three more pending. Dr. Garzon is the president of the ElectrochemicalSociety.

Eric Brosha is a staff member in the Materials Physics and Applications division(MPA-11) at the Los Alamos National Laboratory. He received his B.A. (Summa CumLaude) in Physics from Rider College, Lawrenceville, NJ in 1989. He was awardedan Ashton Fellowship from the University of Pennsylvania, Philadelphia, PA, in1989 and received his Ph.D. in Materials Engineering in August 1993. Currently, hisresearch interests include synthesis of PEM fuel cell catalysts, electrochemical gassensors, materials chemistry and electrochemistry of high-temperature solid-oxidefuel cell electrolyte and electrode materials, X-ray diffraction, X-ray fluorescencespectroscopy and thermal analysis of materials.

Alexandre V. Morozov was born in Rostov-on-Don, Russia. He received his Ph.D. inPhysics from the University of Washington in 2003. From 2003 to 2007 he was apostdoctoral associate at the Center for Studies in Physics and Biology, Rockefeller

at the BioMaPS Institute for Quantitative Biology. In 2009, he was a recipient of anAlfred P. Sloan Research Fellowship. His current research interests include biologicalphysics, non-equilibrium dynamics, and Bayesian modeling.