Biomimetic Olfactory Chips Based on Large-Scale Monolithically Integrated Nanotube Sensor Arrays
Biomimetic Olfactory Chips
Abstract
This paper introduces biomimetic olfactory chips (BOCs) that integrate nanotube sensor arrays on nanoporous substrates, achieving up to 10,000 individually addressable sensors per chip. These chips utilize an engineered material composition gradient and artificial intelligence to offer high sensitivity and distinguishability for various gases and odors. The BOCs can be combined with vision sensors on a robot dog to identify objects in a blind box, showcasing their potential in advanced robotics.
Introduction to Artificial Olfaction and E-Noses
Artificial olfaction and electronic noses (e-noses) aim to replicate the biological olfactory system for discerning complex odorant mixtures. Existing commercial e-noses include AerNos, Alpha MOS Technologies, Sensigent, Aryballe Technologies, and Cyrano Sciences. Current e-nose systems face challenges such as:
Difficulty in miniaturization
Poor recognition capabilities for complex odorant mixtures
Inability to determine exact gas species and concentrations
Addressing these issues is crucial for applications in:
Advanced robots
Portable smart devices
Security patrols
Rescue operations
Mammals detect odors using olfactory receptors and neural networks. The human olfactory system (Fig. 1a–d) involves:
Odor/gas molecules dissolving in mucus on the olfactory epithelium.
Binding to olfactory receptors on the cilia of olfactory neurons, instigating an electrical response.
Transmission of the electrical response through olfactory nerves and fibers to the olfactory bulbs for processing.
Comprehensive processing by the olfactory cortex to discriminate odors.
Canines, with approximately 1,000 types of olfactory receptors, have superior olfaction compared to primates (Fig. 1e).
A smart chemical-sensing unit, typically a gas sensor array or a monolithically integrated chip, determines the performance of an e-nose system. Compared to biological systems, these sensor arrays have a restricted range of ‘olfactory receptors’.
Mainstream gas sensors include:
Metal oxide (MOX) sensors
Polymer sensors
Surface acoustic wave sensors
Quartz crystal microbalance sensors
Electrochemical sensors
Nondispersive infrared sensors
These sensors often have limitations such as large form factors, high power consumption, poor selectivity, and narrow selection bands. MEMS technology allows monolithic integration of millions of identical sensors, but creating diversity beyond hundreds on a single chip is challenging.
## Challenges in Constructing Diversified Sensor Arrays
Constructing diversified sensor arrays involves methods like drop-coating, printing, chemical vapor deposition, or physical vapor deposition. These methods have drawbacks:
Drop-coating or printing: poor mechanical strength
Chemical vapor deposition or physical vapor deposition: inferior sensing performance due to small surface area-to-volume ratio
Techniques like glancing angle deposition have been used to create nano-rod nanostructures with enhanced sensing properties. However, ensuring consistent reproducibility for mass production remains challenging.
Creating hundreds of different gas-sensing materials is difficult, and most sensors use similar materials with different dopants. Devices like the Kamina electronic nose, which use a temperature gradient and a gradient of overlapping membrane layers, face challenges such as high power consumption and limited diversity. Current algorithms in e-nose systems are also relatively singular and difficult to adapt to different applications.
Biomimetic Olfactory Chips (BOCs)
This article reports on biomimetic olfactory chips (BOCs) based on large monolithically integrated sensor arrays (Fig. 1f,g). Each BOC includes:
A high-density array of vertical MOX nanotubes with a well-defined three-dimensional (3D) nanostructure
A range of pixel types
A peripheral signal read-out circuit
Advanced neural network algorithms
The number of sensors in a BOC ranges from 100 to 10,000. The monolithic 3D chemical sensor-array chip is fabricated using atomic layer deposition (ALD) of MOX and subsequent suspended mask-assisted sputtering (SMAS) of diverse MOX.
Fabrication and Structure
The BOC device structure involves a vertical sensor structure with a nanoscale sensing-material layer on a nanotubular porous alumina membrane (PAM), enabling excellent detection sensitivity to ppb-level concentrations of molecular gas. The SMAS method forms a multi-component interfacial (MCI) layer with a gradient distribution of MOX elements on the free-standing PAM.
The device structure includes:
MCI layer on top of the PAM substrate
$PdO/SnO_2$ nanotube sensing-material layer in the PAM channel
Top and bottom sensing electrodes
Insulating layer
Pt heater
The fabrication method is not CMOS compatible (Extended Data Fig. 1). A 10 × 10 sensor array has individual sensors of size 140 × 140 µm2, encompassing an area of about 8 . The pores are about 300 nm in diameter, providing a large surface-to-volume ratio.
A dense insulating $SiO_2$ layer with a thickness of approximately 2.5 µm delivers excellent electrical insulation between the signal electrode and the underlying Pt heating electrode. The Pt heater is crucial for elevating the temperature of the 3D sensor-array chip.
Multi-Component Interfacial (MCI) Layer
The manufacturing process produces pixels with a wide range of responses within a single 3D sensor-array chip using multi-step SMAS to construct an MCI layer. This layer governs the transport of carriers between the top electrode and the sensing-material layer.
The multi-step SMAS method creates a two-dimensional (2D) MOX composition gradient in the MCI layer by successively depositing four different MOXs (ZnO, NiO, $In2O3$, and $WO3$) onto the PAM surface after the ALD of the $Pd/SnO2$ nanotube sensing-material layer.
The spatial distribution of elements in the MCI layer is measured using time-of-flight secondary ion mass spectrometry (ToF-SIMS). The SIMS depth profiles are acquired in 16 regions with a 4 × 4 array. Each of the four cations ($Zn^+$, $Ni^+$, $In^+$, and $W^+$) has a gradient distribution along the four orthogonal directions in 2D space.
New compounds, including $ZnxWO3$ (monoclinic, JCPDS no. 43-1035), $NiWO4$ (monoclinic, JCPDS no. 51-225), $In{2.2}WO3$ (hexagonal, JCPDS no. 37-30) and $In2O3(ZnO){17}$ (rhombohedral, JCPDS no. 43-621), are formed during postannealing at 450 °C.
The sensing-material layer in the PAM channels has a key role in determining the final performance of the 3D sensor-array chip. ALD is employed to achieve the uniform deposition of materials in the PAM channels. $SnO_2$ and Pd layers are successively deposited and annealed at 450 °C to improve their crystallinity. The uniform distribution of elemental Sn and Pd in a single PAM channel is verified by TEM and EDS mapping.
The crystal structure and composition of the sensing-material layer were determined by XRD, XPS, and high-resolution scanning TEM. The tetragonal structures of $SnO2$ (JCPDS no. 41-1445) and PdO (JCPDS no. 41-1107) were confirmed, forming the $PdO/SnO2$ heterojunction sensing-material layer.
Device Measurement and Consistency Validation
A peripheral read-out circuit solves the sneaky path issue induced by the crossbar structure of the electrodes. The resistance of each pixel in the monolithic 3D sensor-array chip can be read out accurately. A 10 × 10 sensor-array chip was used as a proof of concept, suspended and bonded onto a dual in-line ceramic package. At a working temperature of 175 °C, the power consumption was about 240 mW. The coefficient of variation (CV) of resistances is 48.79%. The statistical distribution of the resistances of pure $SnO2$ (CV = 7.12%) and $PdO/SnO2$ (CV = 7.98%) sensor-array chips without the MCI layer was also evaluated.
Twelve 100-pixel BOCs were prepared and tested to investigate device-to-device repeatability. Ten sensors at fixed positions were selected. The maximum CVs for the resistance and gas response were 17.7% and 17.6%, respectively.
Eight gas species (acetone, carbon monoxide, ethanol, formaldehyde, nitrogen dioxide, toluene, hydrogen, and isobutylene) with concentrations ranging from 20 ppb to 4 ppm were tested. Pearson correlation matrix illustrates a discernible correlation between adjacent pixels in the sensor-array chip. A gradual increase in diversity is observed with increasing distance from the initial pixel.
The dynamic sensing response of ten sensing pixels along the diagonal line of the BOC to 1 ppm to 100 ppb acetone is shown in Fig. 3c. Some sensors have a drifting baseline and relatively long response/recovery times due to the relatively low working temperature (175 °C). Commercial MOX sensors are usually operated at 300–400 °C. The BOC has an excellent response to ppb-level gas concentrations (Supplementary Table 4).
The BOC was systematically tested at different temperatures (125, 175, and 250 °C). The response/recovery times gradually decreased with an increase in the working temperature. The gas response reached a maximum at 175 °C. $175 °C$ was chosen as the optimal working temperature.
Response values were normalized to a greyscale between 0 and 255, choosing the responses to 5 ppm $H_2$ as the ceiling value. The normalized response patterns were highly consistent over a range of concentrations and humidity levels. Artificial intelligence algorithms were developed to support the BOC by processing the response patterns.
Single-Gas Discrimination with the BOC System
An end-to-end convolutional neural network (CNN) was employed to classify the gas from the response patterns. A misclassification originates mainly from the low concentrations of gas species with similar functional groups, such as acetone and formaldehyde. For the entire dataset (4,160 samples) of eight gases, the optimized model reached a prediction accuracy of 99.04%.
Sensing data from the first month were used as the training dataset, and sensing data from the second and third months were used as testing datasets. The prediction accuracies in the second and third months were 98.55% and 96.67%, respectively.
Scale of Sensor Array Versus Gas Classification Accuracy
Experiments were designed to illustrate the role of increasing the variety of sensors in achieving higher accuracy. A prediction accuracy of approximately 80% was achieved using only four sensors. Gases with different humidity levels were considered as different species, resulting in a total of 37 species. Accuracy improved rapidly from 15.98% (single sensor) to 93.83% (100 sensors) as the number of sensors used for classification increased.
Using more sensors yields higher accuracy. However, if the number and diversity of pixels were increased, the more complicated reading out of the signal and data redundancy would have to be taken into account.
Deconvolution of a Gas Mixture with the BOC System
A quantitative analysis of gas mixtures with two of four typical gases (ethanol, toluene, formaldehyde, and carbon monoxide) in different ratios was conducted. The dataset included 96 kinds of mixtures with concentrations ranging from 200 to 800 ppb. The response patterns changed smoothly with increasing concentration.
The gas response patterns were first dimensionally reduced to ten-dimensional vectors by principal component analysis. Subsequently, a five-layer fully connected neural network was trained end to end to predict the concentrations of two-gas mixtures. The model has four output neurons to represent the concentrations of ethanol, toluene, formaldehyde, and carbon monoxide.
The average relative error between the predicted concentration and the actual concentration of mixtures was calculated as:
,
where $PregasN$ is the average of the predicted concentrations of gas N and $ActgasN$ is the actual concentration of gas N. The maximum ΔR increased from 7.05% to 8.12% for 200 ppb of a gas species between the data for the second and third months. For the highest concentration (800 ppb), ΔR was lower than 3%.
Real-Time Odour Classification with the BOC System
24 objects with different odours were selected for odour classification. t-distributed stochastic neighbour embedding was employed to convert the 100-dimensional Euclidean distance into 2D coordinates. A support vector machine algorithm with a linear kernel was adopted to divide up the decision space. The BOC system demonstrates excellent classification capability.
The capability of the BOC system to identify the freshness or safety of food is also validated with a time-dependent test of a sliced orange.
Fusion of Olfactory and Vision Sensors on a Robot
A BOC system was installed on a robot dog to bestow the olfactory function on the robot. Five similar boxes were placed in a line, two of which contained objects with different odours (a sample of red wine or a piece of an orange). The robot recognizes the shape of a box with its vision and identifies the internal objects (red wine, a piece of orange, or empty) utilizing its olfaction capability. This demonstrates the immense potential of robots equipped with olfaction capabilities for a diverse array of future applications.
Construction of Large Sensor Arrays
A 20 × 20 sensor-array BOC (with 400 sensors) and a 100 × 100 sensor-array BOC (with 10,000 sensors) were fabricated. The 100 × 100 sensor-array chip was fabricated by photolithography, resulting in individual sensors of size 10 × 10 µm2. The 10,000 sensors occupy an area of about 4 . Its gas response patterns to 1 ppm acetone, ethanol, toluene, and formaldehyde demonstrate its ultra-high resolution.
Conclusions
A biomimetic olfactory system based on large monolithically integrated nanotube sensor arrays was reported. The system uses a monolithic 3D MOX sensor-array chip, fabricated on a nanoporous PAM with ALD and a multi-step SMAS method. The MCI layer has a 2D material composition gradient, enabling a large diversity of sensing pixels. Large sensor arrays ranging from 100 pixels to 10,000 pixels can be fabricated. The sensors have ppb-level sensing performance.
With the assistance of different algorithms, the system exhibits high accuracy (up to 99.04%) in classifying several gas species and accurately identifying the components and concentrations of gas mixtures with a maximum absolute relative error of 8.12% during a three-month testing period. It can distinguish 24 odours. The BOC was integrated into a quadrupedal mobile robot, combining visual and olfactory senses.
Limitations include long-term drift of the baseline, relatively high resistance, relatively high power consumption, and a long response/recovery time. Enhancing the manufacturing consistency of the sensor chips is necessary.
Methods
Deposition of a $PdO/SnO_2$ Sensing Film
The device was built using a custom MEMS fabrication process. PAMs were purchased from Shenzhen Top Membranes Technology Co., Ltd. A shadow mask with a 4 mm by 4 mm window size tightly covered the PAM substrate to define the deposition region. An $SnO2$ film was grown by ALD on the PAM substrate using tetrakis (dimethylamino) tin ($C8H{24}N4Sn$, TDMAS, 99.999%) and $H2O$ as the precursors and $N2$ as the carrier and purging gas. The temperatures of the TDMAS and $H2O$ were maintained at 70 and 25 °C, respectively. The temperature of reaction chamber was kept at 150 °C. $SnO2$ was deposited during 60 cycles of the process. Then Pd was deposited during six cycles, also using ALD, with palladium (II) hexafluoroacetylacetone (99.999%) and hydrazine anhydrous as the precursors at temperatures of 80 and 25 °C, respectively. The chamber temperature was 200 °C.
Deposition of a Suspended Mask-Assisted Gradient Film
A mask with a square window measuring 8 mm × 8 mm was suspended over the PAM substrate. A height of 6 mm was selected to deposit four kinds of MOX in the sequence ZnO, NiO, $In2O3$ and $WO_3$. There was a 2 mm shift between the suspended mask window and the mask window on the PAM substrate. The suspended mask was turned through 90° for the deposition of each MOX. The sample was annealed at 450 °C for 3 h following the sputtering process.
Fabrication of Electrodes, Insulating Layer, and Heater
Thermal evaporation was utilized twice to deposit 150 nm of Au on the top and bottom of the sample as the contact electrodes. For the 10 × 10 sensor-array chip, a shadow mask with ten lines (width and gap both 140 µm) was designed to define the electrode pattern. For the 20 × 20 sensor-array chip, the line width and gap of the shadow mask were reduced to 100 µm. For the 100 × 100 sensor-array chip, photolithography with two line widths (10 µm and 25 µm) was performed. After the electrode deposition, a 2.5 µm $SiO_2$ layer was deposited on the bottom side of the sample by electron-beam evaporation as the insulating layer. A 200 nm serpentiform Pt heater was deposited onto the insulating layer by electron-beam evaporation.
Material Characterization
The XRD patterns were collected by an X’pert Pro diffractometer (PANalytical). The SEM images of the surface and cross sections were captured with the JEOL JSM-7100F and JSM-7800F instruments. The TEM image of the single PAM channel was acquired on a JEOL JEM 2010F instrument. The EDS elemental mapping was obtained by SEM and TEM attachments. The XPS measurements were conducted on a multi-technique surface analysis system (Kratos Axis Ultra DLD Spectrometer). The SIMS depth profiles were acquired by a ToF-SIMS spectrometer (Type V, ION-TOF GmbH). The thickness of the film produced by sputtering was characterized by a surface profiler (Alpha-Step D500 stylus profiler).
Sensing Measurements
The sensor-array chips were bonded to a dual in-line ceramic package or leadless chip carrier using copper wire and silver paste. The packaged chip was integrated with an electrical circuit for reading out the electrical signal. The sensing measurements were performed in a 15 × 15 × 10 chamber using a dynamic gas flow method with a homemade gas supply system. The carrier gas was synthetic air composed of 20% oxygen and 80% nitrogen. The testing gas concentration was controlled through the dilution method with a MATLAB program. The gas response is defined as $(Ra − Rg)/Ra × 100\%$ (reducing gases) and $(Rg − Ra)/Rg × 100\%$ (oxidizing gases), where $Ra$ and $Rg$ are the resistances of sensors exposed to background gases and target gases, respectively.
Description of Read-Out Circuit
The resistance measurement of a single pixel was based on a voltage divider circuit with a reference resistor. Feedback was applied to idle electrodes to eliminate cross-talk. The active pixel was selected by connecting the corresponding column electrode to the reference resistor through the first multiplexer (MUX1) while its row electrode was grounded through MUX2. The signal from the voltage divider circuit went through a low-pass filter and was sampled by a 16 bit analogue-to-digital converter (AD7683). The measurement range was 0.02 to 1,000 MΩ. A microprogrammed control unit was used to synchronize these MUXs and the analogue-to-digital conversion sampling.