Artificial Neurons on Flexible Substrates: A Fully Printed Approach for Neuromorphic Sensing

Abstract

Printed electronic devices have demonstrated their applicability in complex electronic circuits. There is recent progress in the realization of neuromorphic computing systems (NCSs) to implement basic synaptic functions using solution-processed materials. However, a fully printed neuron is yet to be realised. We demonstrate a fully printed artificial neuromorphic circuit on flexible polyimide (PI) substrate. Characteristic features of individual components of the printed system were guided by the software training of the NCS. The printing process employs graphene ink for passive structures and In2O3 as active material to print a two-input artificial neuron on PI. To ensure a small area footprint, the thickness of graphene film is tuned to target a resistance and to obtain conductors or resistors. The sheet resistance of the graphene film annealed at 300 °C can be adjusted between 200 Ω and 500 kΩ depending on the number of printed layers. The fully printed devices withstand a minimum of 2% tensile strain for at least 200 cycles of applied stress without any crack formation. The area usage of the printed two-input neuron is 16.25 mm2, with a power consumption of 37.7 mW, a propagation delay of 1 s, and a voltage supply of 2 V, which renders the device a promising candidate for future applications in smart wearable sensors.

Keywords: artificial neural networks; flexible and functional inks; neuromorphic sensing and computing; printed electronics.

Figure 1

Topology of printed ANN with the proposed neuron design, two layers, and a topology of $3\times 2$.

Figure 2

(a) Top and (b) side views of the printed transistor. (c) Top-view SEM image of the transistor structure. Inset shows SEM and AFM micrographs of the interface between graphene film and In${}_2$O${}_3$. Average graphene flake size was about 200 nm. Atomic force micrography shows that the graphene ink had higher roughness of more than 200 nm as compared to that of the indium oxide film, which had a roughness of 6 nm.

Figure 3

(a) XRD diffraction pattern of indium oxide obtained from aqueous nitrate precursors. Films obtained at 300 °C had characteristic peaks of indium oxide and did not show any unwanted peaks. (b) Variation in graphene sheet resistance with layer thickness showed inverse proportionality. Inset shows a photograph of a printed graphene film on 5 × 5 mm${}^2$ glass substrate annealed at 300 °C as used for resistance dependance measurement.

Figure 4

Transfer characteristics of electrolyte gated transistor before and after 200 bending cycles showed merely negligible difference between curves. Inset shows devices on the polyimide substrate bent to a 4.5 mm radius.

Figure 5

Neuron fabrication steps using the printing equipment described in Section 2. For each step, the top and side views of the printed structure are provided. Polyimide was preheated to 400 °C. In Step 1, graphene of varied thickness was printed, reflected in magnitude of resistance postannealing. In Step 2, aqueous indium nitrate was printed, and the substrate was annealed at 300 °C. In Step 3, CSPE was printed across the gate and In${}_2$O${}_3$.

Figure 6

Hardware prototype of fully printed two-input neuron: (a) Circuit schematic; (b) annotated optical micrograph of the printed two-input neuron on flexible polyimide film. The image was digitally postprocessed and stitched together from multiple microscopy photos.

Figure 7

(a) DC measurement of EGT-based diode ($I_{DS}\left(V_{DS}\right)$) and activation function ($V_{out}\left(V_x\right)$). (b) Transient measurements of the printed neuron.