Sniffing speeds up chemical detection by controlling airflows near sensors

Methods . Elephant sniffing and YouTube sound analysis . Sound recordings of a 35yearold female African Elephant ( Loxodonia africana ) of mass 3360?kg and height 2.6?m were taken at Zoo Atlanta in the fall of 2018. We conducted experiments indoors at the edge of the elephant’s enclosure in the mornings before the zoo opened to the public. All experiments were guided by the staff at Zoo Atlanta without any direct contact by the authors. A subdivided 30?cm?×?60?cm padded box was placed at a location ~1?m outside the enclosure where the elephant could not visually see the box due to obstruction by the bars of the enclosure. For each trial, bran cubes were placed at a different location every time, in position inside or outside the box. The curators instructed the elephant to reach for and find the food. The elephant employed multiple strategies to find the food including sweeping its trunk side to side in the box as well as sniffing for the food. In the trials where the elephant used predominately sniffing to search, the inhalations and exhalations were recorded with a Blue Yetiseries Snowball microphone, similar to methods used with dogs 58 . The most distinct audio waveform was produced in a trial where the food was placed behind a circular cutout just smaller than the elephant trunk tip diameter. Out of 20 trials, three sound recordings were clear enough for the sniffing bouts to be distinguished from the sound of the trunk knocking into the walls of the box. The sound recordings of each sniffing bout, including the elephant experiments and nonelephant third party YouTube videos, were manipulated using Audacity’s noise reduction effect to reduce the background noise by ~20?dB. The maximum number of peaks in the amplitude per second corresponding to an audible sniff was used to calculate the sniffing frequency. Videos of a horse, giraffe, rat, and dog were analyzed using this method. The maximum sniffing frequency of the rat and dog were confirmed with points from the literature 9 , 13 , 13 , 19 . All experiments were performed in accordance with relevant guidelines and regulations. The Georgia Tech Institutional Animal Care and Use Committee approved protocol number A18068 entitled, “Elephant Sniffing, Breathing, and Suction” for dates November 12, 2018–November 12, 2019. Gaseous Recognition Oscillatory Machine Integrating Technology (GROMIT) . We designed and fabricated a sniffing device named the Gaseous Recognition Oscillatory Machine Integrating Technology (GROMIT) which mimics the sniffing mechanics of mammals. The device is designed in modular sections for maximum adaptability. The sections include a custom 3D printed PLA plastic diaphragm pump with a rubber membrane, a Sensirion Venturi flow meter, a custom 3D printed PLA plastic sample housing, and four printed circuit boards with 4 Figaro TGS 2610 sensors on each board. A schematic of the device can be seen in Fig.? 4 . A sniff begins with commands from an Arduino Uno microcontroller to a motor controller which in turn sends commands to an Anaheim Automation 15Y2025LW4 stepper motor. The motor’s axial motion is converted from rotational to linear actuation using a custom slidercrank mechanism which is the driving force behind a 3D printed diaphragm pump. The diaphragm pump is shaped in a way to generate the same amount of volumetric flow rate per actuation, thereby mimicking the ability of a mammal’s lungs to expand and contract. By conservation of mass, the relationship between the desired air velocity and system geometry is $$\delta {V}_{{\rm{b}}}=\frac{{U}_{\max }{A}_{{\rm{t}}}}{f},$$ (5) where A t is the crosssectional area of the tubing, δ V b is the volume change of the bellows, \({U}_{\max }\) is the desired maximum air velocity, and f is the desired frequency of sniffing between 0.1 and 10?Hz. The bellows volume and tubing area were designed so that the Womersley number of the flow could be modified between 0.5 and 7.5 to represent almost the full range found in mammals. On the other side of the pump is a Sensirion Venturi flow meter which tracks and verifies the input flow oscillations. The flow meter confirms a onetoone correspondence between the flow velocity and the input motor signal. The flow sensor is also used to ensure the same average flow rate is obtained for each trial. Next is a series of Figaro metal oxide sensors that were powered on at least 1 week before its first use in order to heat up and remove contaminants, a process called burn in. The sensor section incorporates 4 TGS sensors per board in series for a total of 16 sensors. The board draws ~700–800?mA and is kept powered on before?and?after each measurement test to elude transient unstable response that appears when power is applied to the sensor when it was unpowered for some time 59 , 60 . In our experiments, we run the sensors at a voltage of 5.6?V. Last is a section for the test sample to be placed where the headspace is in series with the flow. In order to avoid unwanted dead spaces in the flow path, the tubing cross sectional area was designed to be constant across all sections. The dead volume in the tubing is estimated to be on the order of 60?mL. We mixed our odor source before conducting trials. We performed experiments with three concentrations: 1 part ethanol to 10 parts DI water by volume, equal parts ethanol and DI water, and pure ethanol solutions. To determine the concentration, Henry’s law \(C=\frac{{\rho }_{{\rm{e}}}}{uH{P}_{{\rm{atm}}}}\) was utilized with ethanol density ρ e ?=?789?kg?m ?3 , ethanol atomic mass u ?=?0.04607?kg?mol ?1 , Henry’s constant H ?=?1.9?mol?m ?3 ?Pa ?1 , and pressure P atm ?=?1?atm?=?1,01,325?Pa 34 . Using this conversion, the concentration levels tested were 8.9 to 89 parts per thousand. These reported concentrations are estimated at the inlet and provide upper limits for the expected lower concentrations which make it to the sensors themselves. Flow Visualization . Tracer particles were generated in the form of humid air generated by a Crane humidifier model number EE5301. The humid air was introduced into the entrance of the flow using a tee junction to ensure no net momentum was added to the oscillating flow pattern of the sniffing device, Fig.? 3 . A rectangular channel was built with approximately?the same cross sectional dimensions as the rest of the tubing in order to maintain unidirectional flow. A section of the channel was removed and replaced with an optically clear acrylic section with toothpaste applied to the inside to prevent fogging. A Viper laser model number 370108 by GLD Products was positioned to shine through the top of the channel, illuminating the particles in the middle of the flow. A Phantom Miro model 320s high speed camera with a Canon 65?mm lens recorded the flow for nine total experiments, at three frequencies of 0.3, 1.3, 2.3?Hz, and at three positions (top, middle, and bottom of the channel). Once recording was finished, analysis was done using the Matlab tool PIVLab. We wrote a Matlab script to separate each video into individual frames and convert each frame to greyscale to speed up the PIVLab process. A region of interest was established and each frame was processed in PIVLab before analysis. Stills from video showing the bottom of the channel when sniffing at 0.3 and 2.3?Hz can be seen in Fig.? 5 e, f respectively. Simulations . The flow simulations are conducted using COMSOL Multiphysics in two dimensions. The chamber is represented as a rectangle with dimensions 30?cm?×?2?cm. The entrance and exit regions are 1?cm?×?2?cm rectangles to represent the tubing connected to the test chamber. The inlet condition is set as an oscillating normal inflow velocity with magnitude varying sinusoidally according to the input frequency of the trial. The outlet condition is zero atmospheric pressure. The walls of the chamber are set as no slip boundary conditions. Initially, the air in the chamber is at rest. The system utilizes the default normal sized physicscontrolled mesh. Sniffing scaling models . Here, we present four models for the relationship between sniffing frequency and body size. We begin with a model that consider’s the air’s inertia. Sniff volumes, also known as a sniffing tidal volume, from literature 9 , 13 , 61 follow the trend \({V}_{{\rm{sn}}}=2.15{M}^{0.99}\) ?mL ( N ?=?7) where M is body mass in kg. Using this scaling, a 20kg dog inhales 42?mL of air during each sniff cycle, the same volume as a shot glass. For comparison, a mouse inhales 0.045?mL of air each sniff, the same volume as an eyedropper drop, and an elephant inhales 4.6?L, the same volume as 1.2 gallon jugs. The control volume \({V}_{{\rm{sn}}}\) in Fig.? 2 b denotes the sniffing tidal volume before it is inhaled into the lungs. The lung volume is generally 25 times larger than the sniff volume, as shown by Stahl’s measurements of lung volume, V lung ?=?53.5 M 1.06 ?mL ( N ?=?333) 24 . When an animal inhales, it uses its diaphragm to apply a pressure P to an airway with a cross sectional area \(\pi {r}_{{\rm{t}}}^{2}\) where r t is trachea hydraulic radius, which has been found in experiments by Tenney 62 to scale as r t ?=?0.0023 M 0.4 (with r t in m and M in kg). The force applied to the air may be written $${F}_{{\rm{L}}}=P\pi {r}_{{\rm{t}}}^{2},$$ (6) where we neglect any losses due to viscosity. The maximum pressure \({P}_{\max }\) of the lungs, generated by muscular contraction, is independent of body size, and has constant peak magnitude of 10?kPa 63 . Throughout the duration of the sniff, the pressure is assumed to vary from positive to negative 10?kPa in a sinusoidal fashion. Therefore, the positive and negative mean values of the pressure waveform are \(P=\pm\! \frac{2{P}_{\max }}{\pi }\) 64 . This pressure is sufficiently low that we can consider air to be incompressible. With air being incompressible, we consider a volume V tot of air that must shift in order to accommodate a new sniffing volume \({V}_{{\rm{sn}}}\) to enter the respiratory system, denoted by the short dashed blue and long dashed green lines in Fig.? 2 b respectively. The mass m of the volume may be written as the product of the air density ρ and the total volume, V tot . The total volume of air in the respiratory system V tot may be written as the sum of the vital capacity V c ?=?56.7 M 1.03 ?mL ( N ?=?315) 24 and the functional residual capacity V r ?=?24.1 M 1.13 ?mL ( N ?=?261) 24 : V tot ?=? V c ?+? V r . We approximate this sum using a power law best fit of these two trends, which yields, V tot ?=?83 M 1.06 ?mL. During a sniff, each air molecule is shifted by a distance \(L={V}_{{\rm{sn}}}/(\pi {r}_{{\rm{t}}}^{2})\) during each period 1/ f . Assuming a sinusodial motion of the air with displacement \(s(t)=L\sin (2\pi ft)\) yields an acceleration a ?=? s ″ of magnitude \(4\pi {V}_{{\rm{sn}}}{f}^{2}/({r}_{{\rm{t}}}^{2})\) . By Newton’s second law, the inertial force on the air may be written F a ?=? m a where a is as above and m ?=? ρ V tot . Together, $${F}_{{\rm{a}}}=\frac{4\pi {f}^{2}\rho {V}_{{\rm{sn}}}{V}_{{\rm{tot}}}}{{r}_{{\rm{t}}}^{2}}.$$ (7) The inertial force on the air F a equals the applied force of the lungs F L , given in Equation ( 6 ). Solving for the frequency f yields our theoretical prediction for sniffing frequency which we call f 1 : $${f}_{1}=\sqrt{\frac{P{r}_{{\rm{t}}}^{4}}{4\rho {V}_{{\rm{sn}}}{V}_{{\rm{tot}}}}}.$$ (8) We proceed by substituting scaling power laws for r t , \({V}_{{\rm{sn}}}\) , and V tot ?=? V c ?+? V r into the above equation, which yields the sniffing frequency, $${f}_{1}=17{M}^{0.25}.$$ (9) Our prediction f 1 is almost twice as high as the experimental data which shows mammals sniff at a frequency slower than their physical limits, possibly because it is too taxing on muscles to consistently operate at their maximum rate 65 . We also give a few caveats with regards to the assumption of sinusoidal pressure. Measurements indicate that larger animals maintain isometric scaling of sniffing volumes. However, according to previous work, as Womersley number increases, the volume flow rate decreases due to viscous effects 18 . Thus, larger animals may be applying larger pressures to compensate. This correction would bring our prediction closer to the experimental trend. Previous studies of breathing have shown that breathing is not in fact sinusoidal. For instance, at its natural breathing rate of 2?Hz, a mouse will inhale and exhale within the first 200?ms and then remain still until the next breath. On the other hand, a mouse exploring its environment with a sniffing frequency of 10?Hz will be moving the air for almost the full duration of the sniff 66 . Since the goal is to create a simple firstorder model, we do not attempt to capture these behavioral effects, and continue with assuming a sinusoidal pressure profile. In fact nonsinusoidal sniffing patterns are more difficult to study mathematically, but they give important rationale for the use of our GROMIT device. Since the motor is controlled by a computer, future workers may input different pressure profiles to find their benefits to sniffing. We next present a model of the natural frequency of the respiratory system first proposed by David Leith 22 in 1983. Starting in the 1960s, breathing and panting were modeled by considering the chest cavity as a damped springmass oscillator. This model is based on experimental measurements of lung compliance C , resistance R , and inertance I on humans and anesthetized animals. For regular respiration, the time scale of respiration relies on the respiratory system’s resistance and compliance. On the other hand, for high speed sniffing, resistance is negligible compared to inertance. The resulting system has qualities similar to oscillating systems such as the forearm muscle 67 , and electrical circuits 50 . Here, the predicted sniffing frequency f 2 can be expressed as $${f}_{2}=\frac{1}{2\pi {(CI)}^{1/2}},$$ (10) where compliance C may be written C ?=?1.59?×?10 ?5 M 1.04 ?L?Pa ?1 where M is in kg, based on N ?=?114 mammals 24 . Inertance of the respiratory system is dominated by the inertance of the air in the trachea, as in our previous model. Based on the idea that inertial pressures should be invariant with body size, Leith 22 , 68 proposed inertance I ?~? M ?1/2 . Using this exponent, the prefactor can be estimated from Spells 50 , who gathered N ?=?15 humans, dogs, and cats from previous workers across a decade of body mass. Using Fig. 7 of Spells’ work, we extrapolate the data points to find I ?=?7.84 M ?1/2 ?Pa?L ?1 ?s ?2 . Combining these power laws, Leith’s prediction yields $${f}_{2}=14{M}^{0.25},$$ (11) which corresponds to the blue dashed line in Fig.? 2 a. It is noteworthy that Leith’s theoretical model has the same exponent as our first model Eq. ( 9 ) and a very similar prefactor (14 vs 17). Furthermore, each model relies on independent measurements: the first model f 1 relies on geometrical meaurements, and Leith’s model f 2 relies on pressure measurements. Their agreement suggests a consistent physical picture. Overall, these models suggests that sniffing frequency aligns with the respiratory system’s natural frequency. Previously, panting was also proposed to correspond to natural frequency 49 . In our next model, we give an upper bound for the sniffing frequency and address the role of viscosity, which has not been considered in the previous two models. Dissipation by viscosity is expected to be important for two distinct flow regimes in the airways 69 . In the regime of slow air flows, slower than the regular breathing rate, inertial effects are reduced and viscous dissipation dominates. The regime of fast air flows is also potentially dissipative due to the generation of turbulent flow structures and their subsequent energy cascade down to dissipative lengthscales. In order to estimate the occurrence of turbulent flow structures in sniffing, we must take into account both the Womersley number Wo and the Reynolds number Re of the flow in the airways 23 , 70 . For the regime Wo? ? ?1, pulsatile flows have their viscous effects confined to a Stokes boundary layer much thinner than the airways diameter 70 . However, the Womersley number of sniffing animals is at most of the order of the unity 18 and thus we instead apply a standard criterion of critical Reynolds number to determine the threshold to turbulence. The maximum Reynolds number associated with laminar flow in the airways is: $${{\rm{Re}}}_{\max }=\frac{2{U}_{\max }{r}_{{\rm{t}}}}{\nu },$$ (12) where \({U}_{\max }\) is the maximal velocity of the displaced volume of air, the trachea radius 62 r t ?=?0.0023 M 0.4 , with M in kg and r t in m, and ν ?=?1.48?×?10 ?5 ?m 2 ?s ?1 is the kinematic viscosity of air. Evaluating the maximal velocity as that of the moving plug of air, \({U}_{\max } \sim 2\pi f{V}_{{\rm{sn}}}/(\pi {r}_{{\rm{t}}}^{2})\) where \({V}_{{\rm{sn}}}=2.15{M}^{0.99}\) ?mL is the sniffing volume, leads to a relationship between the sniffing frequency and maximal Reynolds number $${f}_{\max }=\frac{\nu {{\rm{Re}}}_{\max }{r}_{{\rm{t}}}}{4{V}_{{\rm{sn}}}}.$$ (13) As shown by the experiments of Winter and Nerem in 1984, turbulence 23 in the airways will be unlikely if \({{\rm{Re}}}_{\max }\, <\, 2000\) , which can be therefore written in terms of maximal sniffing frequency: \({f}_{3}\, <\, {f}_{\max }\) so that $${f}_{3}\, <\, 7.91{M}^{0.60},$$ (14) with M in kg and f 3 in Hz. As shown in Fig.? 2 , the blue long dotted line is above all observed animal sniffing frequencies except for animals of mass larger than 30?kg, such as the horse, giraffe, and elephant. This model strengthens our confidence in neglecting viscous dissipation in our originally proposed model for smaller mammals, and simply balancing air inertia and lung force. Lastly, we present a model based on work by Loudon and Tordesillas 18 , who sought to characterize unsteady flow situations similar to those experienced during a sniff. In their model, the amplitude of an oscillating volume flow rate, Q , is related to the maximum pressure P , the radius of the channel r t , the kinematic viscosity μ , and the Womersley number Wo by the equation $$Q\approx \frac{2P{r}_{{\rm{t}}}^{3}}{\mu {\text{Wo}}^{2}}.$$ (15) Using a flow rate approximated as the sniff frequency times the total volume of air in the respiratory system Q ?=? f V tot and Womersley number according to equation ( 2 ), equation ( 15 ) can be solved for frequency f 4 to be $${f}_{4}=\sqrt{\frac{P{r}_{{\rm{t}}}}{\rho \pi {V}_{{\rm{tot}}}}}.$$ (16) Evaluating equation ( 16 ) with a pressure P of 10?kPa 63 , a trachea radius r t ?=?0.0023 M 0.4 ?m 62 , air density, and respiratory volume V tot ?=?83 M 1.06 ?mL produces a frequency $${f}_{4}=470{M}^{0.34}$$ (17) in Hz as shown as a blue dotdashed line of Fig.? 2 a. This trend line is more than an order of magnitude above the experimental data, indicating that Loudon’s assumption of an infinite channel does not wellmatch the finite channel of the trachea. .

From：


