3. Conclusion
We have demonstrated an artificial multisensory integration neuron with
haptic and temperature perception behaviors based on VO2volatile memristor coupled with piezoresistive sensor. Such spiking
neurons can be used to sense different pressure inputs and convert them
into spike trains as a result of the voltage dividing effect between the
piezoresistive sensor and VO2 memristor. Besides, the
spiking neuron is also capable of sensing temperature, by taking
advantage of the intrinsic thermal sensitivity of metal-insulator
transition in VO2. Such spiking neuron is utilized to
recognize Braille characters based on integration of multiple spatial
correlated sensory stimuli, and the coordination of haptic and
temperature sensory inputs give rise to recognition of multimodal
haptic/temperature patterns. It should be noted that the
VO2 based sensory neurons exhibit fast speed, acceptable
cycle-to-cycle and device-to-device variations, but still show
relatively high power consumption, owing to the relatively highV th (~1.4 V) and the low
resistance of LRS (~500 Ω). It is expected that theV th could be reduced by decreasing the length of
the channel of the
VO2 memristor (see
detailed results in Figure S15, Supporting Information), and the
on-state resistance of the device might be reduced by further optimizing
the growth of the VO2 films. Our work offers new
insights into neuromorphic perceptions and neuromorphic computing and
the memristor based multisensory component shows great potential in
cyborg systems, humanoid robotic systems, human–machine systems, and
prosthetic system.
4. Experimental Section
Fabrication of VO2 volatile memristor : The 20 nm
VO2 films were epitaxially grown on
c-Al2O3 substrates by pulsed-laser
deposition (PLD) technique using a 308-nm XeCl excimer laser operated at
an energy density of about 1 J/cm2 and a repetition
rate of 3 Hz. The VO2 films were deposited at 530 °C in
a flowing oxygen atmosphere at the oxygen pressure of 2.0 Pa. Then, the
films were cooled down to the room temperature at the speed of 20
°C/min. The deposition rate of VO2 thin films was
calibrated by X-ray Reflection (XRR). All VO2 devices
studied in this work were fabricated on
Al2O3 substrates. First,
~10 nm single crystal VO2 film serving
as the switching layer was deposited by pulsed laser deposition (PLD).
Afterwards, ~5 nm thick Ti was deposited as the
electrodes and capped by ~40 nm thick Au protection
layer by e-beam evaporation, where the patterning of the electrodes was
done by electron beam lithography and lift-off processes.
Microstructural and compositional characterization : The TEM
samples in this work were prepared by the focused ion beam (FIB)
technique using a dual-beam FIB system (FEI Helios Nanolab workstation).
During FIB patterning, the sample was first coated by
SiO2 and Pt layer deposited using the electron beam to
avoid surface damages, followed by higher-rate Pt coating using normal
ion beam process that served as majority of the protective layer during
FIB cutting. TEM and STEM images as well as EDS measurements were
performed on FEI Tecnai F20 and the HRTEM and SAED results were analyzed
by the Digital Micrograph software (Gatan Inc.). The SEM
characterization was conducted on a field emission SEM (Merlin Compact).
Electrical measurements : All the electrical measurements were
performed using an Agilent B1500A semiconductor parameter analyzer and
the RIGOL MSO8104 digital storage oscilloscope. Voltage pulses were
applied by the Agilent B1500A. We used an Agilent B1500A semiconductor
parameter analyzer to perform electrical measurements of a single
VO2 device in Fig. 2g-h, 4b, 6b
and Figure S2-6, S9, S11, S12. In
Figs. 3-6 and Figure S14, Agilent B1500A is applied to create the pulse
signal, and one channel of the oscilloscope is used to measure the
output of Agilent B1500A, while the other channel measures the voltage
of the output node in the neuron circuit.
Simulations : A COMSOL Multiphysics package finite element
simulation was used to analyze the electro-induced phase change and
thermal distribution of the film. The 3D structure and material
properties of the simulation model, including test electrodes,
substrates, and VO2 composite films, which were the same
as those of the experiment. In this simulation, the fitted electrical
characteristics of the device is set according to the real test data
(see Figure 2f).
The Multi-layer-perceptron with an architecture of 2×20×11 and 3×20×11
were simulated in MATLAB, that is, 2 or 3 input neurons for dataset
inputs, 20 hidden neurons and 11 output neurons for possible classes.
Each output neuron represents one of the combinations of pressure and
temperature. The neural network was trained online with Backpropagation
(BP) algorithm.
Piezoresistive sensor : The
piezoresistive sensors used in this study is RP-C18.3-LT.
The
detailed technical and physical properties are as follows: