Artificial Olfactory Neuron for an In‐Sensor Neuromorphic Nose

Abstract A neuromorphic module of an electronic nose (E‐nose) is demonstrated by hybridizing a chemoresistive gas sensor made of a semiconductor metal oxide (SMO) and a single transistor neuron (1T‐neuron) made of a metal‐oxide‐semiconductor field‐effect transistor (MOSFET). By mimicking a biological olfactory neuron, it simultaneously detects a gas and encoded spike signals for in‐sensor neuromorphic functioning. It identifies an odor source by analyzing the complicated mixed signals using a spiking neural network (SNN). The proposed E‐nose does not require conversion circuits, which are essential for processing the sensory signals between the sensor array and processors in the conventional bulky E‐nose. In addition, they do not have to include a central processing unit (CPU) and memory, which are required for von Neumann computing. The spike transmission of the biological olfactory system, which is known to be the main factor for reducing power consumption, is realized with the SNN for power savings compared to the conventional E‐nose with a deep neural network (DNN). Therefore, the proposed neuromorphic E‐nose is promising for application to Internet of Things (IoT), which demands a highly scalable and energy‐efficient system. As a practical example, it is employed as an electronic sommelier by classifying different types of wines.


Supplementary text 1. Spiking frequency (f) equation of the artificial olfactory neuron module
The relation between input current (I in ) and the output voltage (V out ) of the 1T-neuron can be expressed with the following equation: where C par is the parasitic capacitance connected in parallel to the 1T-neuron.By using the resistance of the SMO gas sensor (R SMO ) to represent I in , the spiking frequency (f) of the artificial olfactory neuron module can be expressed using the following equations: , where V top is the top voltage of V out , V bottom is the bottom voltage of V out , and V DD is the operating voltage applied to the SMO gas sensor.As a result, it was confirmed that f is inversely proportional to R SMO .

Power consumption of the artificial olfactory neuron module
Two kinds of power consumption were extracted for the artificial olfactory neuron module.
The first is the peak power consumption (P peak ), which is the power consumed at the moment of firing.Thus, P peak is simply estimated by the product of I peak V DD , where I peak is the peak current.From the measured I peak shown in Figure S8, P peak of 165 μW was extracted for the artificial olfactory neuron module composed of the SnO 2 gas sensor and the 1T-neuron at NH 3 of 0.5 ppm.The second is the average power consumption (P avg ), which is the average power consumed in one spiking cycle.It can be extracted by integrating the current using the following equation: P avg =  ∫  .As a result, P avg of 350 nW was extracted.

Supplementary figures
Figure S1.Fabrication procedure for a SMO gas sensor.a) A SiO 2 layer of 1 μm thickness was deposited on a Si wafer using the plasma-enhanced chemical vapor deposition (PECVD)     The source meter supplies power to the microheater underneath the SMO gas sensor.The parameter analyzer applies voltage to the SMO gas sensor (variable resistor) and the 1Tneuron, and measures the output voltage (V out ). (V out -t) of the artificial olfactory neuron module when a V G of 2 V was applied for inhibition.
The measurement was performed with the artificial olfactory neuron module composed of the SnO 2 gas sensor and the 1T-neuron.Regardless of the CO concentrations, neuronal spiking was inhibited.In a biological olfactory system, lateral inhibition of the mitral cell controlled by interneurons in the olfactory bulb is important for adaptation and signal contrast.Responses to a) Shiraz and b) Merlot.Shaded and unshaded areas represent time intervals exposed to the wine gas and to the air, respectively.The total gas flow rate injected into the gas chamber was set to 500 sccm, and the wine gases were injected with a flow rate from 1 sccm to 4 sccm.Both wine gases have reducing gas properties.while the other artificial olfactory neuron module was composed of the WO 3 gas sensor and the other 1T-neuron.Capacitors were connected to control the spiking frequency.The capacitor connected to the SnO 2 gas sensor was larger so as to decrease the spiking frequency, because it was found that the responsivity of the SnO 2 gas sensor to the wines was larger than the responsivity of the WO 3 gas sensor.The voltages applied to the sensor (V DD,sen ) and to the gate of the 1T-neuron (V G ) were set at 7.5 V and 0 V, respectively.The synapses had a 1T1R structure, which comprised a commercial stand-alone MOSFET and a single-typed resistor.
The high weight synapse had resistance of 10 Ω and the low weight synapse had resistance of 10 kΩ.The voltages applied to the resistor in the 1T1R synapse (V DD,syn ) and to the source of by E-beam evaporation.The Au decorated metal oxide films were patterned by lift-off, and thermal annealing was performed in N 2 at 400°C for 2 hours.f) The Si substrate located under the microheater was etched away by XeF 2 vapor etching.

Figure S2 .
Figure S2.Fabrication procedure for a 1T-neuron.a) A p-type (100) SOI wafer with a buried-oxide (BOX) thickness of 140 nm was used as a starting wafer.b) The top Si was thinned down to 50 nm and the active area was patterned to define the channel width (W), using photo-lithography and plasma etching.Body doping with boron implantation (1.4×10 13 cm -3 , 10 keV) and rapid thermal annealing (1000 °C, 5 sec) were then carried out.c) A gate dielectric with an equivalent oxide thickness (EOT) of 13 nm and a gate electrode of in-situ doped n + poly-Si with an 100 nm thickness were stacked.The gate area was then patterned by photo-lithography and plasma etching to define the gate length (L G ). d) Source/drain doping

Figure S3 .
Figure S3.Inhibitory function of 1T-neuron.a) Output characteristics (I D -V D ) of the 1Tneuron depending on applied gate voltage (V G ).When a V G of 2 V was applied, single transistor latch (STL) was disabled due to large current flowing regardless of the drain voltage (V D ). b) Spiking characteristics (V out -t) depending on the applied V G .Neuronal spiking was inhibited for a V G of 2 V. C) Spiking frequency (f) versus input current (I in ) depending on the applied V G .When a V G of 2 V was applied, f was extracted as zero regardless of I in owing to inhibited firing.

Figure S4 .
Figure S4.Measurement setup for artificial olfactory neuron module.The measurement apparatus for the artificial olfactory neuron module was composed of gas chambers (air, NH 3 , CO, acetone, NO 2 ), a mass flow controller (MFC) to control the gas concentrations, a source meter, a parameter analyzer, and a probe box to enclose the artificial olfactory neuron module.

Figure S5 .Figure S6 .
FigureS5.Gas independent spiking characteristics of disconnected 1T-neuron.Ratio of spiking frequency in a gas environment (f gas ) and spiking frequency in air environment (f air ) depending on the gas concentration of a) NH 3 b) CO c) acetone d) NO 2 .For all gases, f gas /f air was not varied.Thus, the introduced gases into the probe box did not affect the MOSFET characteristics in the 1T-neuron.

Figure S7 .Figure S8 .
Figure S7.Power consumption of the artificial olfactory neuron module.In order to determine the power consumption of the artificial olfactory neuron module, outgoing source current (I S ) from the source of the 1T-neuron was measured.The measurement was performed with the artificial olfactory neuron module composed of the SnO 2 gas sensor and the 1Tneuron, in a gas environment of NH 3 (0.5 ppm).Extracted P peak was 165 μW when firing occurred.There was no power consumption during the integration process.

Figure S9 .
Figure S9.Dynamic responses of SnO 2 and WO 3 gas sensors to the wine gases.

Figure S10 .
Figure S10.Circuit diagram of the E-nose hardware used for wine classification.The circuit includes two artificial olfactory neuron modules that act as input neurons.One artificial olfactory neuron module was composed of the SnO 2 gas sensor and one 1T-neuron, the 1T1R synapse (V S,syn ) were 5 V and 3 V, respectively.The output currents from the synapses (I syn ) were measured to compare the spiking frequency of I syn1 and I syn2 .

Figure S11 .
Figure S11.Lateral inhibition for wine classification.The synapse current collected at output layer ① (I syn1 ) and at output layer ② (I syn2 ) when the lateral inhibition was applied.a) When the wine was 'Merlot', the artificial olfactory neuron module composed of WO 3 and the 1T-neuron was inhibited by applying a V G of 2 V. b) When the wine was 'Shiraz', the artificial olfactory neuron module comprising SnO 2 and the 1T-neuron was inhibited.Such lateral inhibition can enhance signal contrast and the energy efficiency by firing of only a specific neuron.