Waveguide-Integrated Mid-Infrared Photodetection using Graphene on a Scalable

Chalcogenide Glass Platform

Jordan Goldstein, Hongtao Lin,∗ Skylar Deckoff-Jones, Marek Hempel,
Ang-Yu Lu, Tomás Palacios, Jing Kong, Juejun Hu, and Dirk Englund†
Massachusetts Institute of Technology, 77 Massachusetts Ave, Cambridge MA 02139, USA

Kathleen A. Richardson
University of Central Florida, 4000 Central Florida Blvd, Orlando, FL 32816
(Dated: January 3, 2022)
The development of compact and fieldable mid-infrared (mid-IR) spectroscopy devices represents
a critical challenge for distributed sensing with applications from gas leak detection to environmental
arXiv:2112.14857v1 [physics.optics] 29 Dec 2021

monitoring. Recent work has focused on mid-IR photonic integrated circuit (PIC) sensing platforms
and waveguide-integrated mid-IR light sources and detectors based on semiconductors such as PbTe,
black phosphorus and tellurene. However, material bandgaps and reliance on SiO2 substrates limit
operation to wavelengths λ . 4 µm. Here we overcome these challenges with a chalcogenide glass-
on-CaF2 PIC architecture incorporating split-gate photothermoelectric graphene photodetectors.
Our design extends operation to λ = 5.2 µm with a Johnson noise-limited noise-equivalent power
of 1.1 nW/Hz1/2 , no fall-off in photoresponse up to f = 1 MHz, and a predicted 3-dB bandwidth
of f3dB > 1 GHz. This mid-IR PIC platform readily extends to longer wavelengths and opens the
door to applications from distributed gas sensing and portable dual comb spectroscopy to weather-
resilient free space optical communications.

Here we exceed the wavelength limit of previous

demonstrations using graphene-based detectors on an
Mid-IR absorption spectroscopy is a critical tool for extended-transparency waveguide platform. While
chemical sensing and analysis, especially for inert gases graphene integrated detectors have shown promise at
that evade detection by chemical reaction-based sensors. telecom wavelengths[10], the material’s advantages are
Many such gases derive their inertness from halogenated magnified further at longer wavelengths due to the ther-
chemistries and thus exhibit global warming potential mal nature of the photothermoelectric (PTE) response
due to carbon-halogen stretching modes resonant in the mechanism[11, 12] and due to the impact of optical
thermal IR[1, 2]. To facilitate sensor deployment for plasmon scattering at short wavelengths[13]. Integrated
greenhouse gas leak detection and other chemical sensor photodetection with graphene has been demonstrated
application areas, there exists a strong need to transition at wavelengths up to 3.8 µm[6] and with chalcogenide
from co-packaged discrete components to compact and glass waveguides[14], but on SiO2 platforms. To access
chip-integrated sensors. longer wavelength operation and achieve good sensitivity
To address this challenge, mid-IR photonic integrated at zero bias, we introduce a Ge28 Sb12 Se60 (GSSe)-on-
circuit platforms have been investigated to reduce opti- CaF2 waveguide platform supporting gated PTE-based
cal gas sensors to the size of a chip. Recent work has graphene photodetectors.
demonstrated integrated optical methane[3] and volatile Figs. 1a and 1b illustrate the platform and photode-
organic compound[4] sensing, but required coupling to tector design. The device consists of a single-mode GSSe
off-chip sources and detectors. However, integrating the waveguide fabricated on top of a 5.4 µm wide by 300 µm
detector on-chip is more compact and can improve sensi- long, CVD-grown graphene channel, flanked on either
tivity by reducing the volume of active material able to side by source and drain contacts. Beneath the graphene
generate thermal noise. Su et al. achieved integration channel are pair of CVD graphene back-gates, separated
of a PbTe photoconductor and demonstrated methane by a 400 nm gap and used to electrostatically induce a p-n
sensing at a wavelength of λ = 3.31 µm[5], but their junction along the center of the channel. We use HfO2 as
platform is limited to λ . 4 µm due to absorption in the gate dielectric and as an airtight capping layer. The
the SiO2 substrate[6] and by PbTe’s absorption cutoff[7]. device is fabricated on a CaF2 substrate, transparent up
Waveguide-integrated detectors based on narrow-gap 2D to λ = 8 µm. Fig. 1c depicts the resulting waveguide
materials black phosphorus[8] and tellurene[9] have also mode at λ = 5.2 µm.
been demonstrated, but they too are bandgap-limited to We use lock-in measurement to characterize our detec-
λ . 4 µm. tors, focusing light from a λ = 5.2 µm QCL source into
our chip’s input facet. Light exiting the chip is focused
onto an InAsSb photodetector and amplified for trans-
∗ Present address: College of Information Science and Electronic mission measurement. Supplementary Fig. 1a depicts
Engineering, Zhejiang University, Hangzhou, China. this beam-path in more detail. We operate the device
† Corresponding author: englund@mit.edu under zero bias voltage to avoid introducing electronic
 2

a) Ge28Sb12Se60 TABLE I: Device parameters and approximate values

Capping dielectric
waveguide
Drain contact τDC Drude scattering time at DC ≈ 3.5 fs
Graphene channel
τIR Drude scattering time at IR 30–50 fs
Source contact Gate dielectric
Standard deviation of native carrier 1.5–2.5 ×
Right gate σn
Left gate concentration due to spatial inhomogeneity 1012 cm−2
contact
contact
y EFc Native Fermi level of graphene channel ≈ 0.17 eV
Graphene EFg Native Fermi level of graphene gates ≈ 0.48 eV
x
CaF2 substrate backgates
z
τeph Electron-phonon cooling time ≈ 50 ps
b) αe Excess light attenuation within device 2–3 mm−1
Source Gates c)

culate the electronic temperature distribution and See-

beck photovoltage in the graphene channel under illumi-
nation. Figs. 3a and 3b compare our measured and mod-
y

elled voltage responsivities using calculations described

100 μm in the Methods section. The performance of our de-
vice depends on several fitting parameters, whose defi-
Waveguide Drain
nitions and approximate values (derived from our mea-
sured data) we provide in Table I. We describe our fitting
FIG. 1: a) Illustration of the device cross-section process in Supplementary Section C. Critically, all fea-
perpendicular to the waveguide axis. The optical mode tures of the modelled responsivity map in Fig. 3b up to
supported by the GSSe waveguide evanscently couples an overall scale factor from τeph are established a priori
to and is absorbed by the graphene channel, which is from fitting parameters extracted from the device trans-
gated by two graphene back-gates to induce a p-n mittance and resistance maps, with only τeph obtained
junction. b) Optical image of the device depicting by matching the scales of the measured and modelled re-
source, drain and gate contact pads. c) Depiction of the sponsivities. The resemblance between Figs. 3a and 3b
optical guided mode at λ = 5.2 µm. thus reflects the validity of our PTE model and is not
due to over-fitting. In Fig. 3c, we plot the solution to
Eqn. 6, ∆Tel (x), as well as the source term Q̇(x). The
shot noise and to prevent channel conductivity fluctua- thermal transport model predicts that 9 µW of guided
tions from manifesting as 1/f noise[15]. For the following power raises the temperature of the graphene channel’s
low-frequency measurements we use a lock-in amplifier to electron gas by as much as 1 K along the center of the
measure the photovoltage directly with no preamplifica- device.
tion. Current modulation of our QCL source permits fre-
Figs. 2a, b and c plot the photovoltage, resistance, and quency response measurements up to its modulation
transmission lock-in signals versus both gate voltages for bandwidth of 1 MHz. To account for the modulation
one such photodetector (“Device A”). Here, we modulate response of our laser, we measure the photovoltage of
the λ = 5.2 µm QCL source at 3.78 kHz with a guided Device A alongside that of a fast InAsSb photodiode.
“on” power of 11 µW at the detector input. From our The comparison shown in Fig. 4a indicates that our de-
photovoltage and resistance maps, alongside the power vice is faster than our laser’s modulation bandwidth. We
and waveguide loss calibrations described in Supplemen- thus use a COMSOL model to find the actual RC con-
tary Section A, we infer the gate voltage pairs that op- tribution to our device’s frequency response, plotted in
timize the voltage responsivity, current responsivity, and the inset of Fig. 4a. We also plot the product of the
NEP with respect to Johnson noise, indicated with green RC-limited frequency response and the τeph -limited fre-
markers in Fig. 2. For these, we arrive at 1.5 V/W, quency response with an assumed (1 + (2πτeph f )2 )−0.5
10. mA/W, and 1.1 nW/Hz1/2 , respectively. The ob- dependence, which p applies as the electron-phonon cool-
served photovoltage gate map indicates a PTE response ing length ` = κτeph /Cel ≈ 230 nm is narrower than
mechanism, evidenced by the six-fold sign change pattern our device channel[12]. We thus predict a 3-dB cutoff
around the graphene channel’s charge neutral point[11]. frequency of f−3dB ≈ 1.3 GHz, dominated by the capac-
Figs. 2d, e and f show line slices of the voltage maps as itance between the graphene back-gates.
indicated by the dashed lines in Figs. 2a, b and c of the To investigate our device’s noise performance, we mod-
same color. Fig. 2d, in particular, highlights the changes ulate the QCL current at 30 kHz, amplify the photovolt-
in slope associated with PTE-based detectors[11]. age with a low-noise preamplifier and inspect using a sig-
To confirm our understanding of device operation and nal analyzer. We observe in Device A no broadening of
elucidate the prospects for performance improvement, we the 30 kHz photoresponse peak at offset frequencies as
apply the formalism introduced in Song et al.[12] to cal- low as 0.1 Hz, indicating long-term responsivity stability,
 3

a) b) c)

★: Maximum voltage responsivity = 1.5 V/W ▲: Maximum current responsivity = 10. mA/W ✚: Minimum NEP (Johnson Noise) = 1.1 nW/Hz½
d) e) f)

FIG. 2: a) Measured zero-bias photovoltage produced by the device as a function of the two gate voltages. b) Total
device resistance as a function of the two gate voltages. c) Lock-in signal reflecting power measured by an InAsSb
photodetector at the focal point of our output facet collection lens, used to monitor transmission of the device as a
function of the gate voltages. The power-normalized transmitted is plotted in Supplementary Fig. 3b. d, e, f) Plots
of line sections indicated with dashed lines in figures a, b, and c, respectively.

and we observe no illumination-dependence of the noise Assuming a measurement bandwidth of 0.1 Hz over which
floor. We then measure the un-illuminated noise spectral we have measured our photoresponse to be stable, we find
density and resistance versus both gate voltages. Fig. 4b pgas,min = 23 ppm, roughly equal to the National Insti-
shows the resulting data for a Device B of identical design tute of Occupational Safety and Health (NIOSH) rec-
to Device A, organized by resistance and compared to the ommended exposure limit (REL) of 25 ppm[18]. Remov-
expected Johnson noise spectral density. We observe ex- ing the slightly lossy HfO2 dielectric underneath the gas-
cellent consistency between the measured and predicted light interaction waveguide could decrease pgas,min con-
noise, with a 2–4 dB discrepancy consistent with the spec- siderably, as waveguide losses down to 0.7 dB/cm have
ified noise figure of our preamplifier, corroborating our been demonstrated at the same wavelength using a sim-
earlier claim of Johnson-noise-limited NEP. ilar chalcogenide glass and liftoff process[19].
To demonstrate our device’s utility, we analyze its pre-
dicted gas-sensing performance, summarized from Sup- Although our demonstration is limited to λ = 5.2 µm
plementary Section F. The minimum detectable gas con- by light source availability, the optical conductivity of
centration for a given waveguide platform and photode- our graphene inferred from the fitting parameters in Ta-
tector is given by[16]: ble I remains relatively constant and even increases at
longer wavelengths due to intraband absorption as shown
αbase NEP in Supplementary Fig. 7. We thus expect our platform
pgas,min = e , (1)
ang ΓE I0 to scale to λ = 10 µm and beyond, perhaps requiring a
BaF2 substrate for extended transparency, with little re-
where I0 is the source power, αbase is the waveguide at- duction in performance owing to the PTE effect’s thermal
tenuation coefficient in the absence of gas, a is the specific nature. In Table II we compare our device’s performance
attenuation coefficient of the gas, ng is the guided mode with various off-the-shelf detectors. Although its NEP
group index, ΓE is the confinement factor of electric field is not yet on par with commercial options, its predicted
energy within the gaseous medium, and e = exp(1). For bandwidth may be useful for dual-comb spectroscopy-
detection of nitric oxide (NO), with an absorption peak at based integrated gas analyzers[22]. Additionally, the vac-
λ = 5.24 µm and a specific attenuation of approximately uum requirement of VOx bolometers may complicate co-
a ≈ 70 m−1 atm−1 at√low concentrations[17], we arrive at packaging and introduce coupling losses, and the high
pgas,min = 74 µatm/ Hz for a 1 mW illumination source. cost of HgCdTe may preclude use in broadly deployed
 4

a) a)
VG1=-2V
VG2=3V

b)
b)
Noise spectral density
averaged from 22–32 kHz

FIG. 4: a) Comparison of the frequency response of our

photodetector with that of the laser current modulation
itself. The consistency between the two indicates that
c) the photodetector frequency response exceeds 1 MHz.
Inset: Simulated GHz-range photodetector frequency
response, with and without considering the impact of
the electron-phonon cooling time τeph . b) Measured
noise spectral density versus resistance and
corresponding Johnson noise spectral density of Device
B, without illumination, for the 49 pairs of gate voltages
{Vg1 , Vg2 } where each Vgn is varied from −6 V to 6 V in
steps of 2 V. Measurement was performed at T = 293 K.

1.3 GHz and we predict similar performance at longer

FIG. 3: a, b) Contour plots of the a) measured and b) wavelengths for scaled-up devices enabled by the trans-
modelled responsivity maps of our device, evaluated parency of GSSe beyond λ = 10 µm[23]. Finally, we have
with τDC = 3.5 fs, τIR = 40 fs, σn = 2 × 1012 cm−2 , shown that our device and waveguide platform would en-
τeph = 50 ps, and αe = 2.5 mm−1 . c) Electron able NO detection at concentrations comparable to its
temperature increase ∆Tel and absorbed optical power REL. Substantial improvements are likely using metal-
per area Q̇ profiles in the graphene channel per guided insulator-metal[10] or dielectric slot waveguides to con-
optical power at gate voltages of {−2.35 V, 0.35 V}, centrate the optical mode to within a cooling length of
chosen to maximize the modelled photoresponse, and the pn-junction, which would also reduce the detector
other parameters as above. length and thus device footprint needed to absorb an op-
tical signal. Gapped bilayer graphene may also be inves-
tigated as an alternative to monolayer graphene to reduce
sensor networks. thermal noise[24]. The PIC platform further promises to
In conclusion, we have demonstrated a PTE-based support a full toolkit of mid-IR active devices including
graphene photodetector, integrated in a scalable chalco- on-chip quantum cascade light sources[25], and may even
genide glass waveguide platform with an NEP of leverage the same graphene material platform for de-
1.1 nW/Hz1/2 and a bandwidth exceeding f−3dB = vices such as graphene modulators[14] and hot-electron-
1 MHz. We have modeled the bandwidth to approach based[26] or gapped bilayer graphene light sources. The
 5

HgCdTe PD HgCdTe PD VOx This ing an Elionix FLS-125 125 keV electron beam lithog-
λopt = 5.0 µm λopt = 10.6 µm bolometer work raphy system to pattern alignment marks on our sub-
NEP
√ 1, λ = 5.2 µm 10, λ = 5.2 µm strate, followed by room-temperature development in 3:1
0.9 1100
[pW/ Hz] 0.2, λ = 5.0 µm 40, λ = 10.6 µm isopropanol:methyl isobutyl ketone, e-beam evaporation
f−3dB 1300
of 5 nm Ti/100 nm Au (Temescal VES2550), and liftoff.
1.3 106 10 Hz To transfer the first layer of graphene, we first coated
[MHz] (pred.)
one side of the CVD graphene-on-Cu sheet with PMMA
Vacuum and removed the graphene from the other side using 90
No No Yes No
required? seconds of oxygen RIE (16 sccm He and 8 sccm O2 at a
Waveguide- pressure of 10 mTorr and an RF power of 100W, “oxy-
No No No Yes
integrated? gen RIE process”). We then etched away the Cu using a
FeCl3 -based etchant, followed by 2 DI water rinses, a 30-
TABLE II: Comparison of our detector with inferred minute clean in 5:1 DI water:HCl 37% in water to reduce
room-temperature performance metrics for two HgCdTe metal ion contamination, and two more DI water rinses.
photodiodes optimized for two different wavelengths After letting the graphene film sit overnight in the final
(from [20]) and a VOx bolometer (from [21]) available evaporating dish of water, we scooped it out with our
off the shelf. For the photodiodes, the NEP is CaF2 substrate, blew N2 on the film to eliminate most
extrapolated from the specified detectivity for a of the trapped water, and then baked the sample at 80◦
detector scaled down to match the size of a for 30 minutes followed by 160◦ for 2 hours. We then
diffraction-limited spot with NA = 0.3, which is the removed the PMMA from the graphene using acetone
acceptance NA of these detectors. For the bolometer, at room temperature, rinsed it in isopropanol and blew
we give the NEP of a single 17 µm × 17 µm bolometer it dry, and baked the sample in N2 for 1 hour to im-
pixel as calculated from the specified noise-equivalent prove adhesion. To pattern the graphene back-gates, we
temperature difference as described in Rogalski[7]. spun on a layer of 950 PMMA A6, patterned the gates in
the Elionix and developed as described previously, etched
away the exposed graphene using “oxygen RIE process”,
platform could also be adapted to alternative mid-IR and removed the PMMA as described previously. We
waveguide platforms, such as suspended Ge, as neces- then repeated the Ti/Au liftoff process described above
sary to reach longer wavelength ranges[27]. This research to define the contacts to the graphene gates, after which
represents the first foray into waveguide-integrated de- we evaporated 1.5 nm Al (Temescal VES2550) as an ALD
tectors operating beyond λ = 4 µm, paving the way to- seed layer, allowed the thin Al layer to oxide in ambi-
wards 2D-material-enabled integrated mid-IR microsys- ent, and deposited 300 cycles ≈ 30 nm of HfO2 ALD
tems for gas sensing, spectroscopy[22] and free-space op- at 200◦ C (Cambridge Nanotech Savannah 200). To de-
tical communications[28]. fine the graphene channel and channel contacts, we per-
formed another graphene transfer as described above and
repeated the subsequent graphene patterning and contact
METHODS metallization steps, followed by another Al seed layer and
150 cycles of HfO2 ALD to protect the graphene chan-
Photodetector Fabrication nel. Finally, to pattern the GSSe waveguides, we coated
the chip with 495 PMMA A11, used the Elionix to define
the waveguides, developed as described previously and
A continuous monolayer graphene film was grown on evaporated 750 nm of Ge28 Sb12 Se60 , followed by a quick
Cu foil (99.8%, Alfa Aesar, annealed, uncoated, item no. liftoff in boiling acetone, IPA rinse and N2 blow-dry, and
46365) cut to a size of 15 × 2 cm2 in a 1-inch-diameter cleaving of the chip to expose waveguide facets.
quartz tube furnace under atmospheric pressure. The
furnace was heated to 1060◦ C over 30 minutes under
500 sccm of Ar flow; afterwards, 15 sccm of H2 and 10
sccm of dilute CH4 (1% in Ar) were introduced as re- Measurement conditions
ducing gas and carbon source, respectively, and flowed
for 4 hours to ensure the continuity of the graphene film. The maps in Figs. 2a, b, and c were measured by
Finally, the furnace was allowed to cool to 100◦ C with- sequentially measuring each data point column by col-
out modifying the gas flow before the CVD graphene was umn, bottom to top from left to right. SR830 lock-in
removed from the chamber. Our devices were fabricated amplifiers were used for all measurements. Prior to each
on a 1” diameter by 1.0 mm thick (111)-cut CaF2 sub- data point collection, both gate voltages were reset to
strate (MTI Corporation, item CFc25D10C2). We first −7 V for 80 ms to reset the gate dielectric hysteresis (see
coated our substrate with a PMMA bilayer for liftoff Supplementary Section B), then set to the desired gate
(495 PMMA A6 followed by 950 PMMA A2), which voltages and allowed to dwell for 200 ms for the lock-in
features a slightly re-entrant sidewall profile after de- signal to stabilize. The lock-in filter was set to a 30 ms
veloping. We then performed e-beam lithography us- time constant with a 12 dB/octave falloff. The detector
 6

photovoltage in Fig. 2a was measured directly by the conductivity σIR affect various intermediate model pa-
lock-in amplifier with no additional amplification. For rameters; σDC and σIR themselves depend strongly on
the resistance map in Fig. 2b, we used our lock-in am- EF , which features spatial variation due to the back-
plifier to bias the device with a 1 VRMS sine wave at gates. For the graphene channel, we assume a constant
3.78 kHz through a 100 kΩ resistor to act as a current Nc = N0,c + e−1 Cg Vg in the region above each gate,
source and measured the voltage across the device with where Nc is the carrier concentration in the channel (pos-
the lock-in. To produce the frequency response plots in itive for positive EF , negative for negative EF ), N0,c is
Fig. 4a, we apply a sinusoid of variable frequency to the native carrier concentration at zero gate voltage, Cg
the current modulation input of our QCL and measure is the capacitance per area of the gate dielectric, and
the calibration and photoresponse signals with a SR844 Vg is the voltage applied to the gate in question. (Us-
RF lock-in amplifier. For the laser modulation response ing a set of test devices, we measure Cg = 34. fF/µm2
(indicated in red in Fig. 4a), we couple the laser light on our chip, corresponding to a back-gate dielectric con-
through a single-mode waveguide on our chip with no stant of K ≈ 12; this is described in more depth in
devices on it and directly measure the amplified trans- Supplementary Section D.) In the part of the graphene
mission signal produced by the fast InAsSb detector on channel above the gap between the two gates, we as-
the output side of our chip. For the photovoltage sig- sume a linear slope between Nc,1 and Nc,2 . For the
nal (blue curve in Fig. 4a), we amplify the photovoltage gates, Ng = N0,g − e−1 Cg Vg , with Ng and N0,g de-
produced by our detector by 40 dB using a preamplifier fined similarly to Nc and N0,c . In general, the graphene’s
and measure this amplified signal with our lock-in. In all Fermiplevel and carrier concentration are related by EF =
cases, we used a dwell time of 1.5 s, and the filter of our ~vgr π|N | sign(N ), where vgr is graphene’s Fermi veloc-
lock-in was set to 100 ms with a 12 dB/octave falloff. To ity. To incorporate the blurring of the graphene’s Fermi
measure the un-illuminated noise spectral density in Fig. level-dependent properties due to spatial carrier concen-
4b, we amplify the noise produced by the device using tration variations, we convolve the Kubo formula with a
a 60 dB preamplifier and analyze the output on an FFT Gaussian as follows:
signal analyzer while controlling the gate voltages applied
to the device. We choose to measure the averaged noise σDC (N ) =
spectral density between 22 and 32 kHz where we find no Z ∞ (n−N )2
1 −1
electromagnetic interference-related spectral peaks. We √ e 2 σn2 σ(0, EF (N ), τDC , T0 ) dn (3)
also use a lock-in amplifier to simultaneously measure the σn 2π −∞
device resistance as described above for Fig. 2b, albeit
at a higher frequency so as to not produce a signal in the and similarly for σIR (N ) using ω = 2πc/λ instead of
−1
noise measurement range. We use our signal analyzer’s 0 and τIR instead of τDC . Finally, we have R = σDC ,
band averaging feature to measure the noise spectral den- κ = π 2 kB
2
T0 σDC /3e2 via the Wiedemann-Franz law, and
sity for each data point. To produce the final plot, we S = −d(log σDC )/dEF [30]. Cel is obtained by convolving
manually record the resistance and noise spectral density the heat capacity of pristine graphene with a Gaussian of
for all gate voltage pairs from −6 V to 6 V in steps of 2 V. standard deviation σN as in Eqn. 3, where the pristine
heat capacity is given by[30, 31]:
 Z ∞
 2|ε| ∂fd (ε − EF (N ))
Device modelling Cel (N ) = ε 2 2 dε. (4)
σn =0 −∞ π~ vgr ∂T
We use the Kubo formula adapted from Hanson[29] to We use a waveguide eigenmode solver to find the mode
model graphene’s conductivity at DC and infrared fre- profile of our waveguide at λ = 5.2 µm, using refractive
quencies (albeit with different values of the Drude scat- indices of 1.4, 2.6, and 1.88 for the CaF2 , GSSe, and
tering time τ for the different frequency ranges): HfO2 , respectively. The resulting mode profile enters into
our expression for Q̇el as follows[32]:
je2 (ω − jτ −1 )

σ(ω, EF , τ, T ) = |Ex (x, yc )|2 + |Ey (x, yc )|2 σIR,c (x)
π~2 Q̇el = P RR . (5)
" Z ∞   R2
Re(E × H∗ ) · ẑ dx dy
1 ∂fd (ε) ∂fd (−ε)
× ε − dε
(ω − jτ −1 )2 0 ∂ε ∂ε Here yc is the y-coordinate of the graphene channel, and
Z ∞ # yg would be the y-coordinate of the graphene gates. We
fd (−ε) − fd (ε) R W/2
− dε (2) may then write αc = P −1 −W/2 Q̇el (x) dx. Similar ex-
0 (ω − jτ −1 )2 − 4(ε/~)2 pressions hold for αg in terms of σIR,g (x), noting of course
that σIR,g (x) = 0 for x within the gap between the gates
where e is the elementary charge, fd (ε) = (exp((ε − R W/2
EF )/kB T ) + 1)−1 is the Fermi-Dirac distribution and where there is no graphene. Finally, ρΩ = −W/2 R(x) dx.
kB is Boltzmann’s constant. As I will show below, Having thus obtained expressions for κ(x), Cel (x),
graphene’s low frequency conductivity σDC and infrared Q̇el (x), S(x), Π(x), αc , αg and ρΩ as a function of the
 7

gate voltages as well as τDC , τIR , σn , EFc , EFg , τeph , αe , coefficient within the detector, including contributions
and ρc , we then solve for the increase in electronic tem- not only from the graphene channel but also from the
perature per guided power ∆Tel (x)/P = (Tel (x) − T0 )/P graphene gates (αg ) as well as a gate-independent excess
using the equation: loss αe associated with scattering and absorption from
organic or metallic impurities attached to or trapped un-
  derneath the graphene sheets. Thus the total device re-
d d ∆Tel −1 dΠ sistance is equal to R = RΩ + RΠ + Rc , and the voltage
− κ + τeph Cel ∆Tel = η Q̇el − Jx , (6)
dx dx dx responsivity is given by:
where κ is the 2D electronic thermal conductivity of
the graphene, τeph is the electron-phonon cooling time, Rv αc

Q̇el is the absorbed optical power per area, η is the Rv = 1 − e−αtot L , (10)
L αtot
conversion efficiency of absorbed optical power to elec-
tronic heat after initial electron-phonon scattering[12], which we plot versus both gate voltages in Fig. 3b for
Jx is the line current density in the x-direction, and Π is the best-fit device parameters given in Table I obtained
the Peltier coefficient. We are approximating the electric as described in Supplementary Section C. All calculations
field to run exclusively in the x-direction, valid for suf- are carried out in Mathematica.
ficiently gradual light absorption. We assume η = 1 as
has been previously reported in pump-probe experiments
at this wavelength range[13]. The thermal electromotive AUTHOR CONTRIBUTIONS
force (EMF) arising from the Seebeck effect is then given
by: J.H., D.E. and J.G. conceived the experiments. J.G.
designed, fabricated, and measured the devices, with
Z W/2 the exception of chalcogenide glass deposition, performed
d ∆Tel
Ex = − S dx, (7) by H.L. and S.D.-J. under the supervision of J.H., and
−W/2 dx graphene growth, performed by M.H. and A.-Y.L. un-
der the supervision of J.K. and T.P. K.R. provided the
where W = 5.4 µm is the channel width and S is the
chalcogenide glass sources for thermal deposition. J.G.
Seebeck coefficient. In Eqns. 6 and 7, κ, Cel , S, and
and D.E. wrote the manuscript. All work was supervised
Π = STel ≈ ST0 (for small ∆Tel ) are all dependent
by D.E.
on the local Fermi level EF of the graphene, and thus
have a gate-tunable x-dependence, which we account for
in our calculations. Combining the equations, the η Q̇el
DATA AVAILABILITY
source term in Eqn. 6 gives rise to a proportional photo-
induced EMF, whereas the Peltier term Jx dΠ dx gives rises
to a current-dependent EMF which appears as a resis- The datasets generated during and/or analysed during
tance in series with the Ohmic and contact resistances of the current study are available in the FigShare repos-
the channel. We can thus write: itory at https://doi.org/10.6084/m9.figshare.c.
5514759.v1.

V = Rv αc P (z) − (ρΩ + ρΠ + ρc ) Jx (z) (8)

CODE AVAILABILITY
where V is the voltage across the contacts, Rv is
the photovoltage per absorbed power per length of a The Mathematica document used to simulate pho-
cross-sectional slice of the device (i.e., dimensions of todetector performance metrics is available in the
V/(W/m)), αc is the component of the waveguide power FigShare repository at https://doi.org/10.6084/m9.
attenuation coefficient arising from absorption in the figshare.c.5514759.v1.
graphene channel, P (z) is the guided power at a posi-
tion along the waveguide, and ρΩ , ρΠ , ρc are the Ohmic,
Peltier, and contact line resistivities (dimensions of Ω·m), ACKNOWLEDGMENTS
respectively. Averaging over z along the length of the
waveguide we obtain:
We would like to thank the MIT.Nano and MIT Nanos-
tructures Laboratory staff for maintaining the cleanroom
Rv αc
facilities used to fabricate these devices, in particular
V = 1 − e−αtot L Pin − (RΩ + RΠ + Rc ) I, (9) Mark Mondol, Jim Daley, and Dave Terry. We also would
L αtot
like to thank Sebastián Castilla of ICFO for helpful dis-
where I is the current produced by the photodetec- cussions. This research was funded in part by a grant
tor, thus describing a Thévenin equivalent source. Here, from the Army Research Office via the MIT Institute for
αtot = αc + αg + αe is the total guided power attenuation Soldier Nanotechnologies University-Affiliated Research
 8

Center (ISN UARC) (award number W911NF-18-2-004), pressed in this material are those of the author(s) and do
the NSF Graduate Research Fellowship Program (award not necessarily reflect the views of the National Science
number 1122374), and NSF Award #2023987. Any opin- Foundation.
ions, findings, and conclusions or recommendations ex-

[1] B. H. Stuart, Infrared Spectroscopy: Fundamentals and tonics 11, 798 (2017).
Applications, 1st ed., Analytical Techniques in the Sci- [15] F. N. Hooge, 1/f noise sources, IEEE Transactions on
ences (Wiley, Hoboken, 2004). electron devices 41, 1926 (1994).
[2] V. Ramanathan, Greenhouse effect due to chlorofluoro- [16] R. Siebert and J. Müller, Infrared integrated optical
carbons: Climatic implications, Science 190, 50 (1975). evanescent field sensor for gas analysis: Part i: Sys-
[3] M. Pi, C. Zheng, R. Bi, H. Zhao, L. Liang, Y. Zhang, tem design, Sensors and Actuators A: Physical 119, 138
Y. Wang, and F. K. Tittel, Design of a mid-infrared (2005).
suspended chalcogenide/silica-on-silicon slot-waveguide [17] E. L. Saier and A. Pozefsky, Quantitative determination
spectroscopic gas sensor with enhanced light-gas inter- of nitric oxide and nitrous oxide by infrared absorption,
action effect, Sensors and Actuators B: Chemical 297, Analytical Chemistry 26, 1079 (1954).
126732 (2019). [18] National Institute for Occupational Safety and
[4] T. Jin, J. Zhou, H.-Y. G. Lin, and P. T. Lin, Mid-infrared Health, Nitric oxide, https://www.cdc.gov/niosh/
chalcogenide waveguides for real-time and nondestructive npg/npgd0448.html, accessed: 2021-03-28.
volatile organic compound detection, Analytical Chem- [19] H. Lin, L. Li, Y. Zou, S. Danto, J. D. Musgraves,
istry 91, 817 (2019). K. Richardson, S. Kozacik, M. Murakowski, D. Prather,
[5] P. Su, Z. Han, D. Kita, P. Becla, H. Lin, S. Deckoff-Jones, P. T. Lin, V. Singh, A. Agarwal, L. C. Kimerling, and
K. Richardson, L. C. Kimerling, J. Hu, and A. Agar- J. Hu, Demonstration of high-Q mid-infrared chalco-
wal, Monolithic on-chip mid-IR methane gas sensor with genide glass-on-silicon resonators, Opt. Lett. 38, 1470
waveguide-integrated detector, Applied Physics Letters (2013).
114, 051103 (2019). [20] Thorlabs, Inc., Mid-IR Photovoltaic Detectors, HgCdTe
[6] A. Yadav and A. M. Agarwal, Integrated photonic ma- (MCT), https://www.thorlabs.com/newgrouppage9.
terials for the mid-infrared, International Journal of Ap- cfm?objectgroup_id=11319, accessed: 2021-03-29.
plied Glass Science 11, 491 (2020). [21] C. C. Li, C.-J. Han, and G. D. Skidmore, Overview of
[7] A. Rogalski, Infrared Detectors, 3rd ed. (CRC Press, DRS uncooled VOx infrared detector development, Op-
Boca Raton, 2019). tical Engineering 50, 1 (2011).
[8] L. Huang, B. Dong, X. Guo, Y. Chang, N. Chen, [22] I. Coddington, N. Newbury, and W. Swann, Dual-comb
X. Huang, W. Liao, C. Zhu, H. Wang, C. Lee, and K.-W. spectroscopy, Optica 3, 414 (2016).
Ang, Waveguide-integrated black phosphorus photode- [23] P. Klocek and L. Colombo, Index of refraction, disper-
tector for mid-infrared applications, ACS Nano 13, 913 sion, bandgap and light scattering in gese and gesbse
(2019). glasses, Journal of Non-Crystalline Solids 93, 1 (1987).
[9] S. Deckoff-Jones, Y. Wang, H. Lin, W. Wu, and J. Hu, [24] M.-H. Kim, J. Yan, R. J. Suess, T. E. Murphy, M. S.
Tellurene: A multifunctional material for midinfrared op- Fuhrer, and H. D. Drew, Photothermal response in dual-
toelectronics, ACS Photonics 6, 1632 (2019). gated bilayer graphene, Phys. Rev. Lett. 110, 247402
[10] P. Ma, Y. Salamin, B. Baeuerle, A. Josten, W. Heni, (2013).
A. Emboras, and J. Leuthold, Plasmonically enhanced [25] C. Tsay, F. Toor, C. F. Gmachl, and C. B. Arnold,
graphene photodetector featuring 100 Gbit/s data recep- Chalcogenide glass waveguides integrated with quantum
tion, high responsivity, and compact size, ACS Photonics cascade lasers for on-chip mid-IR photonic circuits, Opt.
6, 154 (2019). Lett. 35, 3324 (2010).
[11] N. M. Gabor, J. C. W. Song, Q. Ma, N. L. Nair, T. Tay- [26] L. Kim, S. Kim, P. K. Jha, V. W. Brar, and H. A. At-
chatanapat, K. Watanabe, T. Taniguchi, L. S. Levitov, water, Mid-infrared radiative emission from bright hot
and P. Jarillo-Herrero, Hot carrier–assisted intrinsic pho- plasmons in graphene, Nature Materials https://doi.
toresponse in graphene, Science 334, 648 (2011). org/10.1038/s41563-021-00935-2 (2021).
[12] J. C. W. Song, M. S. Rudner, C. M. Marcus, and L. S. [27] A. Osman, M. Nedeljkovic, J. S. Penades, Y. Wu, Z. Qu,
Levitov, Hot carrier transport and photocurrent response A. Z. Khokhar, K. Debnath, and G. Z. Mashanovich, Sus-
in graphene, Nano Letters 11, 4688 (2011). pended low-loss germanium waveguides for the longwave
[13] K. J. Tielrooij, J. C. W. Song, S. A. Jensen, A. Cen- infrared, Opt. Lett. 43, 5997 (2018).
teno, A. Pesquera, A. Zurutuza Elorza, M. Bonn, L. S. [28] A. Soibel, M. W. Wright, W. H. Farr, S. A. Keo, C. J.
Levitov, and F. H. L. Koppens, Photoexcitation cascade Hill, R. Q. Yang, and H. C. Liu, Midinfrared interband
and multiple hot-carrier generation in graphene, Nature cascade laser for free space optical communication, IEEE
Physics 9, 248 (2013). Photonics Technology Letters 22, 121 (2010).
[14] H. Lin, Y. Song, Y. Huang, D. Kita, S. Deckoff-Jones, [29] G. W. Hanson, Dyadic green’s functions and guided sur-
K. Wang, L. Li, J. Li, H. Zheng, Z. Luo, H. Wang, S. No- face waves for a surface conductivity model of graphene,
vak, A. Yadav, C.-C. Huang, R.-J. Shiue, D. Englund, Journal of Applied Physics 103, 064302 (2008).
T. Gu, D. Hewak, K. Richardson, J. Kong, and J. Hu, [30] G. Grosso and G. P. Parravicini, Solid State Physics, 2nd
Chalcogenide glass-on-graphene photonics, Nature Pho- ed. (Academic Press, Oxford, 2014).
 9

[31] A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. [32] A. W. Snyder and J. D. Love, Optical waveguide theory,
Novoselov, and A. K. Geim, The electronic properties of 1st ed. (Chapman and Hall, London, 1983).
graphene, Rev. Mod. Phys. 81, 109 (2009).
 Supplementary Information for Waveguide-Integrated Mid-Infrared
Photodetection using Graphene on a Scalable Chalcogenide Glass
Platform
arXiv:2112.14857v1 [physics.optics] 29 Dec 2021

Jordan A. Goldstein, Hongtao Lin, Skylar Deckoff-Jones, Marek Hempel, Ang-Yu Lu,
Kathleen Richardson, Tomás Palacios, Jing Kong, Juejun Hu and Dirk R. Englund

January 3, 2022

A. Optical Setup, Alignment, and Power and Loss Calibration

a) b)

Chopped λ = 5.2 μm GSSe waveguide To lock-in amplifier

illumination
5 mm offset

InAsSb photodiode
Focusing lens

Iris
Graphene photodetector

CaF2 substrate
Collection lens

Supplementary Figure 1: a) Depiction of the in- and out-coupling beam path for optical and
optoelectronic measurements. Lengths are not to scale; the chip and waveguide are magnified
for clarity. b) Output photodiode signal versus waveguide length for four waveguide kickback
structures. The decaying exponential fit reveals the waveguide loss and the in-coupled power,
subject to a conversion efficiency accounting for the collection efficiency of the collection optics and
the voltage responsivity of the photodiode. Inset: Layout of the kickback structures.

Supplementary Fig. 1a depicts the optical beam path during all optical and optoelectronic
measurements. Chopped, collimated illumination at a wavelength of λ = 5.2 µm is coupled into a
cleaved waveguide facet at the edge of the chip using a molded aspheric focusing lens to achieve a
diffraction-limited spot. The waveguides containing our photodetectors and test devices as well as
the kickback waveguides for loss measurement are designed to exit the chip at a 5 mm offset with
respect to the input to reduce the amount of stray light picked up by the collection optics. The
collection optics consists of a 1” high-NA germanium asphere which images the output facet at
the center of an iris, behind which we place an InAsSb photodiode that monitors the out-coupled
power. To align the setup, we flip the Ge collection lens, iris and photodiode out of output the beam
1
path and use a long focal length CaF2 lens and liquid nitrogen-cooled InAsSb camera to image the
output facet of the chip while adjusting the chip position to achieve coupling first through a straight
multimode waveguide (not shown) followed the desired S-shaped device or kickback waveguide. We
then replace the long focal length lens with the high-NA germanium collection lens, flip the iris
back into the collection beam path in the “open” position, and place the camera behind the iris.
By adjusting the position of the collection lens to focus the out-coupled light at the center of the
iris while gradually reducing the iris aperture size and monitoring the focal point on the camera,
we are able to localize the focus of the collection lens at the center of the iris. Finally, we replace
the camera with the InAsSb photodiode and adjust the in-coupling and out-coupling optics to
maximize the signal measured by the photodiode.
To determine the loss of our waveguides and predict the optical power immediately incident
upon our graphene photodetector during characterization, we have fabricated a set of five waveguide
“kickback” structures of varying length, the layout of which is shown in the inset of Supplementary
Fig. 1b. A large radius of 75 µm is used for the waveguide turns to reduce the associated light
leakage to negligible levels. We individually align and optimize the coupling for each of these
waveguides and record the lock-in signal of the collection photodiode; these data are plotted versus
the total kickback waveguide length in Supplementary Fig. 1b. We could not collect a data point
for the longest kickback because we found the out-coupled light to be indistinguishable from stray
illumination. Fitting the data to a decaying exponential reveals a waveguide loss of 1.1 dB/mm and
an in-coupled power corresponding to a lock-in signal of 46.2 mV. To obtain the actual in-coupled
optical power, we must divide this value by the product of the voltage responsivity of the photodiode
and the efficiency of the collection optics. For the former, we obtain a value of 1.76 × 103 V/W from
the manufacturer-provided calibration data. For the latter, we perform a full-wave electromagnetic
simulation of the waveguide facet and extract the far-field profile of out-coupled light. Integrating
the simulated optical power falling within the entrance pupil of the collection optics, we obtain a
collection efficiency of 0.45. Combining the above three figures, we arrive at 58. mW coupled into
the waveguide facet at optimal alignment. Accounting for waveguide loss, we obtain a power of
11. mW immediately incident upon our measured photodetector, from which we can thus calculate
the responsivity figures reported in the main article.

a) b)
3.2 μm
Ti/Au 750 nm
Ge28Sb12Se60

ALD HfO2 CaF2

Graphene sheets

Supplementary Figure 2: a) Illustration of the test device cross-section perpendicular to the waveg-
uide axis. b) Transmission versus applied voltage for one of our test devices using a zigzag voltage
sweep to reveal hysteresis. Sweep direction is color-coded and indicated with arrows.

2
B. Hysteresis effect

In addition to photodetector devices, we also fabricated gated graphene test devices as illustrated
in Supplementary Fig. 2a. These are similar to the detector devices described in the main article,
except with only a single graphene sheet and a single contact on each graphene layer. Indeed,
the detector devices can be made to behave similarly by applying the same voltage to each of the
back-gates, but we report the test devices to illustrate the gate hysteresis effect observed for all
devices. Supplementary Fig. 2b shows the transmission response versus applied voltage for one of
our test devices. Applying a zigzag voltage sweep reveals an undesirable hysteresis pattern showing
distinct curves for rising and falling voltages sweeps, which we label red and blue respectively. The
same effect is observed in our photodetector devices as well. Hysteresis has been widely reported
for as-deposited HfO2 gates, and is attributed to trapped charge carriers within the dielectric[1, 2].
To prevent the effects of hysteresis from appearing in the gate sweeps presented in Main Figure 2,
we “reset” both gate voltages to 6.5 V prior to collecting each data point.

C. Measured versus modelled resistance and transmittance and parameter ex-

traction

Supplementary Fig. 3 compares the measured and modelled resistance and transmittance gate maps
obtained using the mean graphene quality and waveguide loss parameters listed in Main Table 1;
namely, τDC = 3.5 fs, τIR = 40 fs, σn = 2×1012 cm−2 , and αe = 2.5 mm−1 . We determined the values
of these parameters to achieve the best fit simultaneously between both pairs of maps. Generally
speaking, τDC is inversely proportional to RΩ , τIR affects the scale and modulation contrast of the
transmittance T = e−αtot L , and σN affects the sharpness (width at half-max) and contrast of both.
First, we infer EFc simply from from the gate voltage of the charge neutral point (peak in the case
of RΩ and dip in the case of T ). Since the sharpnesses of both R and T are largely determined
by σN for relatively high σN as is the case for our devices, we then determine σN to best match
both maps. In particular, we find that the resistance map is best fit by σn = 1.5 × 1012 cm−2 and
the transmittance map by σn = 2.5 × 1012 cm−2 , from which we obtain the error margins quoted
for σn in Main Table 1; σn = 2 × 1012 cm−2 represents a compromise between these two values.
We then determine τDC to roughly match the scale of R, but we allow the modelled resistance
to be on the order of 10 Ω less the measured resistance to take into account the possibility of a
contact resistance in this range as justified in Supplementary Section E.. The uncertainty in the
actual contact resistance, as well as the imperfect fit, both contribute to uncertainty in the actual
value of τDC . We determine τIR and EFg to best match the measured transmittance map, and
the values quoted in Main Table 1 represent our best attempt to simultaneously reflect multiple
features of this map; namely, the modulation contrast between the center and corners of the map,
the contrast between the upper right and other corners, and the falloff at the low-voltage edges
of the map (where intraband absorption of the graphene gates is strongest) and the high-voltage
edges of the map (where interband absorption of the graphene gates is strongest). The quoted
error margins of τIR reflect the range over which these different features of the transmittance map
are best rendered in our model, and the error margins of αe reflect the range required to match
the scale of the measured transmittance map over the error range of τIR . We finally use these
six parameters to predict the relative gate dependence of the voltage responsivity. In this way,
only the overall scale factor of the responsivity is subject to a fitting parameter (namely, τeph );
the contour of the responsivity map is, in our device’s performance regime, purely predicted from
parameters extracted from the resistance and transmittance maps. Therefore, the resemblance
between the measured and modelled responsivity maps is not the result of tweaking parameters to
3
 a) b)

c) d)

Supplementary Figure 3: a, b) Contour plots of the measured resistance and transmittance maps
as a function of gate voltages. c, d) Contour plots of the modelled resistance and transmittance
maps as a function of gate voltages using the graphene quality and waveguide loss parameters listed
in Main Table 1.

achieve a fit, but rather reflects the accuracy of the photoresponse mechanism model itself using
model parameters “fed forward” from the resistance and transmittance maps.

D. Measurement of gate capacitance

We used Transmission Line Method (TLM) devices to measure both the gate capacitance and
contact resistance of our devices. Shown in Supplementary Fig. 4, these devices consist of a
large, contacted graphene back gate, gating a set of graphene FETs of increasing channel lengths,
ranging from 5 µm to 65 µm. To extract the capacitance per area of our hafnia dielectric, we
measure the capacitance between the lower graphene gate and the upper graphene/gold structure,
and divide by the overlap area of these two structures. To measure the capacitance, we use a lock-in
amplifier to apply a 0.5 VRMS sinusoid of variable frequency to the rightmost two metal contacts
in Supplementary Fig. 4, and in the current return path we place a 10 kΩ shunt resistor, the voltage
across which we monitor with the lock-in amplifier. At each frequency and for each TLM that we
4
Supplementary Figure 4: Transmission line method (TLM) device used for gate capacitance and
contact resistance measurements.

measure, we perform a measurement with both pads contacted and with only one pad contacted to
compensate for any stray capacitance in our setup. We then subtract the measured capacitances
in the “connected” and “disconnected” cases to obtain the actual device capacitance as a function
of frequency, which we plot in Supplementary Fig. 5. Excluding the orange curve as an outlier, we
measure a capacitance of C ≈ 30 pF corresponding to a capacitance per area of Cg = 3.4 fF/µm2
and a dielectric constant of K ≈ 12.

E. Measurement of TLM structure for gate resistance extraction

In additional to capacitance measurement, we also use the TLM structures for their intended pur-
pose of evaluating the resistance of our graphene-metal contacts. For each of the six channels in
each of the five TLM devices, we measure the channel resistance as a function of gate voltage using
an upward voltage sweep each time to compensate for the hysteresis discussed in Supplementary
Section B.. Since the gate voltage of the charge neutral point may shift slightly between measure-
ments due to trapped charges in the gate, we then shift the measured resistance curves so that
their peaks overlap. Finally, for each measured voltage offset from the Dirac peak and for each
TLM, we fit the data of resistance versus channel length to a line and plot the resulting y-intercept
as a function of voltage offset, manually eliminating any gate voltage sweeps showing malformed
(for instanced, flattened or bimodal) resistance peaks. The resulting y-intercept curves are shown
in Supplementary Fig. 6. Unfortunately, we find highly inconsistent intercept resistances between
the five TLM devices, with the intercept even going negative in several cases. Therefore, we are
unable to draw a quantitative conclusion regarding the contact resistance of our devices. We can
at least, however, estimate the total contact resistance (summing over both contacts) for our TLM
devices to be generally in the ≈ 1 × 102 Ω range; therefore, since the TLM channels are 40 µm wide,
we would expect the total contact resistance of our actual photodetectors to be in the ≈ 1 × 101 Ω
range, which is an order of magnitude lower than our measured resistances; therefore we conclude
that it can be safely ignored in our modelling as other sources of error (such as the imperfect fit
between our measured and modelled resistance and transmittance maps) are much more likely to
dominate the uncertainty in our analysis.

5
 Supplementary Figure 5: Capacitance versus measurement frequency for five devices.

F. Optical mode attenuation due to gas absorption and gas sensitivity analysis

Our gas sensitivity analysis follows that of Siebert at el.[3], which considers a light source at the
gas absorption peak illuminating a long gas-light interaction waveguide, partially cladded with the
ambient air in which gas’s presence is suspected, and terminated in a noisy photodetector. However,
the authors do not give a rigorous analysis of the attenuation coefficient of guided light due to the
absorbing gas, which we supply here. The power attenuation coefficient in a waveguide with a
perturbative source of absorption is given by [4]:
RR 2
2 2nκ |E| dx dy
αwg = ω0 RR R ∗
, (1)
R2 Re(E × H ) · ẑ dx dy
where the complex refractive index n̄ ≡ n − jκ. For a partial pressure of target gas pgas , the free
space attenuation coefficient is:
αfree-space = apgas = 2κω0 /c, (2)
therefore κ = apgas c/(2ω0 ), and thus
RR 2
gas n0 |E| dx dy
αgas = apgas c RR , (3)
R2 Re(E × H∗ ) · ẑ dx dy
where n0 ≈ 1 is the refractive index of the gaseous medium, as κ is only nonzero in the gaseous
region of the waveguide cross-section. Since our waveguide mode is absorbed only very weakly, we
now reference the expressions for group velocity vg and propagation constant β, respectively, given
in Snyder and Love for non-absorbing waveguides[4]:
RR
c2 β R2 Re(E × H∗ ) · ẑ dx dy
vg = RR , (4)
ω0 R2 n2 Re(E × H∗ ) dx dy
and RR
ω0 n2 Re(E × H∗ ) dx dy
R2RR
β= 2 2 2
. (5)
c R2 n |E| dx dy
6
Supplementary Figure 6: Y-intercepts of resistance versus channel length as a function of gate
voltage offset from the resistance peak for our five TLM devices.

Combining these two equations, the group index ng is thus:

RR 2 2
c
RR R2 n |E| dx dy
ng = =c ∗
. (6)
vg R2 Re(E × H ) · ẑ dx dy

By comparison with Eqn. 3, we finally arrive at:

RR 2
gas n0 |E| dx dy
αgas = apgas ng RR 2 |E|2 dx dy
= ang n−1
0 ΓE pgas , (7)
R2 n

where ΓE is the electric field confinement factor:

RR 2 2
gas n0 |E| dx dy
ΓE ≡ RR . (8)
2 2
R2 n |E| dx dy

We emphasize that this is not, in general, equal to the “traditional” confinement factor Γ defined
based on the proportion of modal Poynting vector in gain/loss region, as has been previously
noted[5].
The remainder of our gas sensitivity analysis follows that of Siebert et al.[3], in which an ideal
gas-light interaction wavelength is found as a function of the gas concentration where maximum
sensitivity to deviations is desired. For gas detection applications, maximum sensitivity at zero
concentration is desired, and the ideal interaction waveguide length is simply found to be lopt =
−1
αbase . This then leads directly to Main Eqn. 1.

G. Optical absorption of graphene for longer wavelengths

In Supplementary Figure 7, we apply Main Eqn. 3 to predict how the real part of our graphene’s
optical conductivity (and thus optical absorption) would change at longer wavelengths using carrier
concentration spread and Drude scattering time material parameters extracted from our device
7
Supplementary Figure 7: Contour plot of real part of graphene’s infrared optical conductivity as
a function wavelength and mean carrier concentration, assuming a carrier concentration spread of
σn = 2.0 × 1012 cm−2 and a Drude scattering time of τIR = 40 fs.

model. We find that the optical absorption remains roughly the same if not even higher at longer
wavelengths, which is due to increased intraband absorption; therefore we anticipate similar device
performance can be achieved at these wavelengths.

References

[1] B. H. Lee, L. Kang, R. Nieh, W.-J. Qi, and J. C. Lee, “Thermal stability and electrical char-
acteristics of ultrathin hafnium oxide gate dielectric reoxidized with rapid thermal annealing,”
Applied Physics Letters, vol. 76, no. 14, pp. 1926–1928, 2000.

[2] J. Robertson, “High dielectric constant oxides,” European Physics Journal-Applied Physics,
vol. 28, no. 3, pp. 265–291, 2004.

[3] R. Siebert and J. Müller, “Infrared integrated optical evanescent field sensor for gas analysis:
Part i: System design,” Sensors and Actuators A: Physical, vol. 119, no. 1, pp. 138 – 149, 2005.

[4] A. W. Snyder and J. D. Love, Optical waveguide theory. London: Chapman and Hall, first ed.,
1983.

[5] T. D. Visser, H. Blok, B. Demeulenaere, and D. Lenstra, “Confinement factors and gain in
optical amplifiers,” IEEE Journal of Quantum Electronics, vol. 33, no. 10, pp. 1763–1766, 1997.