CMOS integration of inkjet-printed graphene for humidity sensing

We report on the integration of inkjet-printed graphene with a CMOS micro-electro-mechanical-system (MEMS) microhotplate for humidity sensing. The graphene ink is produced via ultrasonic assisted liquid phase exfoliation in isopropyl alcohol (IPA) using polyvinyl pyrrolidone (PVP) polymer as the stabilizer. We formulate inks with different graphene concentrations, which are then deposited through inkjet printing over predefined interdigitated gold electrodes on a CMOS microhotplate. The graphene flakes form a percolating network to render the resultant graphene-PVP thin film conductive, which varies in presence of humidity due to swelling of the hygroscopic PVP host. When the sensors are exposed to relative humidity ranging from 10–80%, we observe significant changes in resistance with increasing sensitivity from the amount of graphene in the inks. Our sensors show excellent repeatability and stability, over a period of several weeks. The location specific deposition of functional graphene ink onto a low cost CMOS platform has the potential for high volume, economic manufacturing and application as a new generation of miniature, low power humidity sensors for the internet of things.


Microhotplate design and fabrication.
A key component of our resistive humidity sensor is the CMOS MEMS based microhotplate (μHP) structure. Details of the μHP structure are reported elsewhere 33,34 . A cross section view of the sensor structure is shown in Fig. 1(a). An optical microscope image of the fabricated device is shown in Fig. 1(b). The silicon die measures 1 mm × 1 mm. It is designed using a 1.0 μm Silicon on Insulator (SOI) CMOS process technology and fabricated in a commercial foundry, followed by deep reactive ion etching (DRIE) to release the thin membrane. The processing employs SOI wafers with 1 μm buried oxide, 0.25 μm SOI layer and 3 metallization layers. The μHP structure typically consists of an embedded, 0.3 μm thick, resistive tungsten microheater (metal layer 1), 0.3 μm thick heat spreader plate (metal layer 2), and a top gold layer for interdigitated sensing electrodes (IDEs). The diameters of the circular heater and membrane structures are 250 and 600 μm, respectively. The heater is fabricated during the CMOS process while the top gold electrodes and corresponding tracks are deposited as a post CMOS process in the same commercial foundry. The tungsten heater controls the operating temperature of the membrane and graphene-based sensing layer. The IDEs underneath the sensing layer are used to measure the change in resistance due to humidity exposure. Use of tungsten in the heater allows the device to operate at a very high temperature (up to 750 °C), if required, for example, when oxidebased sensing materials are used. Gold is used as the electrode material because of its chemical inertness (and hence, unchanged conductivity over prolonged use under various temperature and humidity conditions) compared to commonly used aluminum in this SOI process. The silicon underneath the μHP is etched away, using the dioxide layer as the etch stopper, at a wafer level by DRIE technique. This forms a 4.5 μm silicon dioxide: SiO 2 (4 μm)/silicon nitride: Si 3 N 4 (0.5 μm) membrane structure onto which the microheater and electrodes are suspended. The membrane structure reduces DC power consumption of the sensing device to < 5 mW when used for humidity sensing in this work. The heating temperature is uniformly confined over the microheater region, due to the buried heat spreader 33 . The temperature decreases rapidly away from the heater region and is at close to room temperature at the membrane rim, allowing reliable temperature independent on-chip circuit performance 35 .
The characteristic power versus temperature plot (up to 217 °C) of the μHP device is given in Fig. 2. To calculate the power consumption, the resistive μHPs are first calibrated to up to 300 °C using a computer controlled high temperature chuck (Signatone S-1060R-6TG). Two temperature coefficients of resistance (α, β) are calculated from the measured value using the relationship: where R 0 is the resistance of the heater at room temperature T 0 , and T is the temperature increase. The values of α (2.05 × 10 −3 K −1 ) and β (0.2 × 10 −6 K −2 ) are very closely matched to the values provided by the CMOS foundry. A constant current is supplied to the µHP and the voltage is measured across the heater. From the constant current and measured voltage, the power supplied to the devices is calculated. The corresponding heater temperature is calculated from the change in resistance using α, β values estimated during the calibration. Over small temperature changes, a linear function is fitted through the experimental points of power versus temperature plot, yielding an electrothermal transduction efficiency of 8.46 °C/mW; Fig. 2.
Formulation of graphene-polymer ink. The first step towards ink formulation is exfoliation of graphite into graphene flakes. UALPE starts with mixing bulk graphite crystals into a solvent. The ultrasound causes high-frequency pressure variations and formation of microcavities in the solvent. Collapse of these microcavities produces high shear forces, exfoliating mono-, bi-and few-layer flakes from the bulk crystals by overcoming the interlayer van der Waals forces 36 . Since the UALPE process does not involve chemical pre-or post-treatment, the dispersed graphene flakes are chemically pristine. Note that, when experimental parameters such as sonication vessel and solvent volume are kept fixed 37 , solvent viscosity plays an important role in cavitation (higher viscosity requires higher acoustic pressure to create cavitation 38 ) and hence, the UALPE process. However, in reality, a large majority of common solvents have a viscosity ranging from 1 to 5 mPa.s at room temperature, which have no strong effect on the exfoliation. Instead, the dispersability and stability of graphene in the solvents are largely governed by their intermolecular interactions. This has been explained through matching the degree of the  It is thus required to develop graphene inks with low surface tension solvents to allow for good wetting. Alcohols, in particular, are attractive for this purpose. In addition, their low boiling point allows rapid drying of ink after deposition, ensuring uniform coating of graphene flakes. Indeed, they are commonly used as the primary or secondary solvent in the majority of graphics and functional inks 44 . Among the common alcohols, isopropyl alcohol (IPA) can exfoliate graphene, but with meta-stable dispersion 45 . This is primarily due to its mismatched Hansen solubility parameters (δ D = 15.8 MPa 1/2 , δ P = 6.1 MPa 1/2 and δ H = 15.4 MPa 1/2 ) with graphene. Therefore, pure IPA based graphene dispersion cannot be used for inkjet printing. UALPE with ionic surfactants and nonionic polymers are commonly used strategies to stabilize graphene 46,47 . This is typically achieved through steric hindrance and favourable enthalpic interactions with graphitic surface 48 . Here, we use PVP as the stabilizer. Indeed, PVP is a polymer analog to NMP with the N-substituted pyrrolidone rings similar to that of NMP 49,50 , and has been used before to stabilize carbon nanostructures, such as nanotubes in different solvents 49,51,52 . In addition, PVP is highly hygroscopic, swelling when absorbing moisture in humid environments 53 . PVP is also an electrically insulating polymer but can be made conductive by addition of graphene through percolation 54 . The hygroscopic nature and moisture-induced swelling with conductivity change in graphene-PVP composites make our formulated graphene-PVP-IPA ink (henceforth, we use the term 'graphene ink') very attractive for humidity sensing.
We ultrasonicate 100 mg of graphite (Sigma-Aldrich, 100 mesh flakes) with 1.5 mg PVP (Sigma-Aldrich, average molecular weight 10,000 Da) in 10 mL IPA for 12 hours at ~15 °C. This starting PVP concentration enables stabilization as well as desired ink rheological properties specific to our inkjet printer nozzle discussed below. The resultant dispersion is centrifuged for 1 hour at 4,030 rpm (~1540 g), sedimenting the unexfoliated graphite. The upper 70% of the dispersion, enriched with mono-and few-layer graphene nanoflakes, is decanted for analysis and experiment.
Optical absorption spectroscopy is used to estimate the concentration of the dispersed graphene using the Beer-Lambert law ( where c is the graphene concentration (gL −1 ) and l is the distance the light passes through the dispersion (m). λ A and α λ are the absorption (a.u.) and material dependent optical absorption coefficient (Lg −1 m −1 ) at wavelength λ (nm), respectively. Figure 3 shows the optical absorption spectrum of the dispersion diluted to 10 vol% to avoid scattering losses during absorption measurement 55 . The inset of Fig. 3 shows a photograph of a cuvette containing the undiluted graphene dispersion. The spectrum is mostly featureless as expected, due to the linear dispersion of Dirac electrons 56,57 . The peak in the UV region is a signature of the van Hove singularity in the graphene density of states 58 . Using α 660 nm = 2460 Lg −1 m −1 47 , we estimate the concentration of graphene in the undiluted dispersion as 0.40 gL −1 . We note that the graphene ink is stable, without forming any visible aggregation over several months.
The graphene flakes are characterized with Atomic Force Microscopy (AFM). The sample for AFM is prepared by dip-coating a Si/SiO 2 wafer into a graphene dispersion diluted to 2 vol% by pure IPA. Though the concentration of PVP in this diluted dispersion is ~3.8 × 10 −4 wt%, the residual polymer prevents accurate measurement of the flake dimensions. The sample is therefore annealed at 400 °C for 30 min. This temperature is chosen because PVP starts to decompose at this temperature in air 59 while the exfoliated graphene flakes remain stable 60,61 . The annealed sample is imaged with a Bruker Dimension Icon AFM in ScanAsyst TM mode, using a silicon cantilever with a Si 3 N 4 tip. Micrographs of typical flakes are shown in Fig. 4(a,b), along with the height variations across the samples; Fig. 4(c,d). The thickness distribution of flakes is also measured; Fig. 4(e). This shows that 63% of the flakes have thicknesses < 5 nm, corresponding to < 13 layers, assuming ~0.7 nm measured thickness for the bottom layer and ~0.35 nm increase in thickness for each additional layer 62 . The average flake lateral dimension is ~204 nm. Note that the number of layers obtained in IPA/PVP is higher than what is typically obtained by UALPE of graphite in water/SDC, with ~26% mono-and ~22% bi-layer, and with 300-600 nm average lateral dimension from similar experimental parameters 47 .
Inkjet printing of graphene. We next consider deposition of our graphene ink to form a graphene-PVP thin film composite as the active sensing layer over the IDEs of the 250 μm diameter CMOS µHP. For this, large area techniques such as spray-, spin-or dip-coating are unsuitable, as they do not offer selective area deposition directly on to the IDEs. We use drop-on-demand inkjet printing technique for the deposition of active materials on to the IDEs. It is a digital printing technique, where single ink droplets are ejected from an ink chamber in response to individual electrical impulses 26 . Inkjet printing also offers selective area deposition across the X-Y plane of a substrate. The inkjet printer used in this work is a DMP-2831 Dimatix Printer. The printing nozzle has a diameter of 22 μm. The volume of individual droplets from this nozzle is ~10 pL. We achieve printing resolutions of ~100 μm with this set-up onto the IDEs without any surface modification. By controlling the number of droplets and the area of deposition, inkjet printing thus enables us to deposit a well defined volume of ink for humidity sensing only onto the IDEs.
A stable drop generation (single droplet generation for each electrical impulse, without the formation of satellite droplets) and jetting of ink (avoiding deviation of droplet trajectory) is of primary importance for high quality inkjet printing. Otherwise, unstable jetting may lead to uncontrolled amount of ink deposition on to undesired locations. In inkjet printing, a figure of merit, Z, is commonly used to consider the printability of inks and is defined as: where γ is surface tension of the ink (mNm −1 ), ρ is the density of ink (gcm −3 ), η is the viscosity of the ink (mPa.s) and a is the nozzle diameter (μm). As a rule of thumb, it is commonly accepted that Z value should be < 14 for the case of drop-on-demand inkjet printing to avoid secondary or satellite droplets 26,63 . Meanwhile, it should be > 1 to optimize droplet formation or ejection to avoid long-lived filament formation, which may result in poor positional accuracy and printing resolution 26,63,64 . The value of Z typically falls within 1-10 for commercial inkjet printable inks 65 . We stress that the acceptable range of Z values should be established by experiments and numerical simulations and should be considered as a guide only. Indeed, inks with values outside this range may also be printable, in particular, by controlling the shape and intensity of the electrical pulses for droplet generation. To determine the Z value of our graphene ink, we use pendant drop measurement and parallel plate rheometry to measure γ and η, respectively at room temperature. We measure γ ~ 28.0 mNm −1 and η ~ 2.34 mPa.s. The density is measured as ρ ~ 0.8 gcm −3 . With a = 22 μm, we calculate Z ~ 9.48, falling into the recommended range for stable jetting. This indicates the formulated ink is suitable for inkjet printing. A stable jetting without the formation of satellite droplet or long filament is thus expected. This is experimentally confirmed by high speed jetting sequence images presented in Fig. 5(a).
The drying process of the deposited ink is critical to ensure uniform coverage after the carrier solvent is evaporated. After impacting the substrate, the wettability of the ink droplets carrying stably dispersed nanoparticles ( . . e g graphene) defines consistency of the particle coating on to the substrate 43 . During the drying process of a droplet, a finite contact edge is formed at the edge of the solid-solvent interface. This enhances the transport of the solvent, promoting faster evaporation at the edges than in the central area of the droplet. Meanwhile, the difference in evaporation speeds results in a convective flow within the droplet from the center towards the edges (Marangoni effect) 66 . This transports the dispersed materials to be deposited at these edges. These deposited dried materials prevent the contact edge from receding, 'pinning' it at its original position. This further promotes the deposition of particles as the solvent evaporates, leading to a non-uniform coating at the edges, commonly termed as the 'coffee ring effect' 67,68 . The low surface tension of our ink ensures a good wetting, and the low boiling point of IPA (82.6 °C) guarantees a quick evaporation at room temperature, leading to a consistent and uniform coverage of graphene flakes across the printed area. Figure 5(c) shows the dark field optical microscope images of the CMOS device with graphene deposited. The bright spots in Fig. 5(c) indicate graphene flakes in the PVP matrix, demonstrating an even deposition of graphene across the IDEs. We also investigate the structure of the deposited graphene-PVP using scanning electron microscopy (SEM); showing evenly interspersed graphene flakes (darker spots) on the CMOS µHP; see Fig. 5(d). Humidity sensing. Evaporation of IPA forms a graphene-PVP composite sensing layer on to the IDEs. Within the graphene-PVP composite, the graphene flakes are randomly distributed. The electrical behaviour of randomly arranged conducting objects in an insulating thin film matrix can be described by percolation theory in 3d 69 . When continuous pathways of the conducting objects are not formed, the conductivity of the matrix is zero. As the conductive pathways start building up, electrical conductivity becomes non-zero and increases with an increasing number of conducting pathways following the relation: where φ is the conductivity, X 0 is the critical density of the conducting objects above which the density of the objects X results in conductivity and i is the percolation coefficient. For a 3d composite with conductive filler materials distributed in an insulating matrix (such as the graphene-PVP composite), the equation can be rewritten in terms of volume fractions: where ν and ν 0 are the volume fraction and the critical volume fraction of the conducting objects, respectively 70 . Since this relation does not take into account of particle size, shape, orientation and their distribution uniformity inside a composite matrix, i and ν 0 are empirically derived for a given system 70 .
Modelling the graphene flakes as thin, circular, 2d conductive platelets and considering their 3d random distribution within the composite, it can be shown that 70 : where d and t are the diameter and thickness of the platelets, respectively, and ⁎ d is the interplatelet distance. If ⁎ d is shorter than the electron hopping distance through the non-conducting matrix, electron hopping takes place, facilitating formation of conductive pathways. The limit for single-step electron hopping is ~10 nm 71 . Multistep electron hopping may take place for > 10 nm. A critical value of ~1 μm has been reported for some polymers with conductive fillers 71,72 . With ~4.5 nm average flake thickness and ~204 nm average lateral dimension estimated from AFM measurements and considering d * ~ 10 nm-1 μm, we get v 0 ~ 0.486-0.003. This requires ν > 0.486 to ensure a conductive graphene-PVP composite. Bulk density of graphite is ~2.3 gcm −3 . However, exfoliated graphene flakes can have a significantly lower density. Commercial graphene samples are typically quoted to have a density ranging from 0.03-0.4 gcm −3 73 . Assuming 0.4 gcm −3 for our UAPLE graphene, a > 0.047 gL −1 concentration of graphene ink is required to form a conductive graphene-PVP composite. We stress that this estimation is dependent on further experimental determination of ⁎ d of PVP and the density of UALPE graphene. The graphene-PVP composite (v ~ 0.889) formed from this graphene ink (c ~ 0.40 gL −1 ) is therefore predicted to be conductive. This is experimentally confirmed with resistance measurement between the IDEs at room temperature, 40 °C and 70 °C, as shown in Fig. 6(a). The linear relationship confirms good Ohmic contact between the graphene-PVP composite and the gold IDEs. A control experiment is conducted by inkjet printing pure PVP solution, which shows negligible electrical conductivity, as expected.
The sensors are next tested using different relative humidity (RH) levels inside a stainless steel chamber. Initially the sensors are kept in dry air before they are exposed to eight different humidity levels (10-80% RH). In typical experiments, the sensors are exposed to humid air for 20 minutes, followed by dry air purging for 30 minutes. Note that our set-up allows a maximum of 80% RH above which condensation of water vapor inside the test chamber starts to appear. The sensor response is defined as: where R humid and R air are the measured resistances exposed to humid conditions and dry air, respectively. Typical response of the sensor with the inkjet-printed graphene-PVP composite sensing layer when exposed to different RH levels at room temperature is plotted in Fig. 6(b). We see an increased response with an increased RH. The sensitivity (defined as: Response/RH) achieved from the sensor varies from 0.3%/%RH to 0.21%/%RH as RH increases from 10% to 80%. We also observe a drift in the response baseline in the lower humidity levels (10-20%). This could be due to trapped water molecules within the sensing layer during the recovery from pre-exposures 74 .
To investigate the effect of temperature on the response, we have tested the sensors at three temperatures (room temperature, 40 °C and 70 °C). The temperature rise is achieved by using the on-chip microheater. The result is plotted in Fig. 6(c). We observe a reduced response with increasing temperature. We note that the baseline drift shown in Fig. 6(b) reduces with increase in temperature as the trapped water molecules desorb from the sensing layer faster. We further add that 40 °C and 70 °C are sufficiently low not to induce any discernible thermal stress on to the CMOS µHP or the graphene-PVP film and affect the measurements.
To further investigate the influence of graphene volume fraction on the sensing layer, a series of 5 inks with 0.40, 0.32, 0.24, 0.16 and 0.08 gL -1 graphene concentration is prepared while keeping the PVP concentration constant. This gives the volume fraction ν as 0.889, 0.865, 0.828, 0.762 and 0.615, respectively. With this v > 0.486, the deposited graphene-PVP composites after evaporation of IPA are predicated to be conductive. Sensors with these inks are then exposed to 40-60% RH. We observe an increased response with the increase of graphene concentration under all the RH conditions; Fig. 6(d). We suggest that the graphene-PVP composite with more percolating pathways shows a larger change in resistance, leading to a higher response. The larger change may be subject to that more percolating pathways are broken due to the polymer swelling effect. However, this is yet to be conclusive, requiring further experimental work to understand the electrical conduction and percolation behavior of our graphene-PVP composite sensing layer.
The sensor with 0.40 gL −1 graphene ink exhibits the highest response and is used to further investigate typical response and recovery times (defined as the time needed to reach 63% of the maximum response and to recover to the baseline, respectively). The response and recovery times measured at room temperature are reasonably fast, varying from ~6-16 s and ~60-300 s, respectively; Fig. 7(a,b). The slower recovery time at higher RH could be attributed to the presence of higher partial pressure of water vapour close to the surface of the sensing layer when the sensor recovers through the desorption of water molecules. It should be noted that long (humidity) ON and OFF times were used to ensure that the device response reaches its saturated limit without any noticeable drift.
Reproducibility of a sensor at different sensing cycles and of sensing over a long period are also investigated. We measure response of different sensors with the same design to investigate performance variations between the devices. All the sensors use 0.40 gL −1 ink and are tested at room temperature. The reproducibility of response is demonstrated at 50% RH for five cycles; Fig. 8(a). The long-term stability of the sensor is investigated over a 4-week period. We do not observe a significant variation (~4%) of response at 30, 40, 50 and 60% RH during this period; Fig. 8(b). The performance variation is investigated by measuring the response for three separate but identical sensors when exposed to different RH conditions, with a maximum variation of ~13%; Fig. 8(c). Figure 8(c) also shows a linear response of the sensors. We believe the variation could be further improved by optimizing the sensing layer and graphene flake dimensions for the development of reliable, reproducible, low power and compact sensors.

Discussion
The adsorbed water molecules on graphene surface may disassociate to H + and OH − ions at the edges of graphene, similar to what is observed for graphene oxide samples 6 . These H + ions may tunnel from one water molecule to another through hydrogen bonding, reducing the overall electrical resistance of the sensing layer 6 . On the other hand, the adsorbed water molecules on graphene surface act as electron donors 5 . As the electron density increases, the normally p-type graphene becomes more resistive. This p-type semiconductor nature of carbon materials is in agreement with reported results 75,76 . Meanwhile, PVP is a hygroscopic, electrically insulating polymer, absorbing up to 25% moisture at 75% RH, swelling the polymer 53 . We propose that moisture absorption and subsequently polymer swelling increases the distance between the graphene flakes, leading to a reduction of electrically percolating pathways through single hopping. We note this change in polymer volume and corresponding increase in distance between the adjacent graphene flakes may also depend on other factors such as chemical composition and graft density of the polymer film 77 . Among the two opposing effects on resistance, we suggest that effect of the p-type semiconductor nature of graphene and the swelling and consequent reduction of percolating pathways dominate. This results in increased electrical resistance of the graphene-PVP composite sensing layer, enabling humidity sensing through the IDEs.
In summary, we have integrated functional graphene inks with CMOS MEMS technology to fabricate a resistive humidity sensor. Our formulation of the ink using a blend of hygroscopic, electrically insulating polymer with chemically pristine, conducting graphene nanoflakes offers an ideal sensing layer for use with an economic, compact and low power CMOS sensing platform. Through 3d percolation theory, we show that our ink formulation produces a conductive sensing layer. Upon humidity exposure, reduction in percolating networks in the sensing layer leads to an increased resistivity for humidity sensing. The response of the sensors are reproducible, with a maximum of ~13% variation among sensors at different RH levels with stable performance (< 4%) over a period of few weeks. The combination of CMOS and inkjet printing platform, with graphene and potentially, other nanomaterial based functional inks, opens exciting opportunities for mass produced CMOS based sensor systems for a tremendous variety of highly commercially valuable applications.