An Inkjet-Printed Artificial Neuron for Physical Reservoir Computing

Inkjet printing circuits onto thin, flexible substrates is a newly explored field with respect to the transistor; a critical element needed to form logic gates and high-level active circuitry. The traditional approach is to achieve comparable performance to MOSFETs or BJTs. However, the introduction of neuromorphics, spintronics, memristors, chaotic materials, and limitations of transistor sizes have incited a shift toward alternative computing schemes that do not behave as standard transistors. This work explores a minimal fabrication, low-cost, nonlinear, current-controlled graphene inkjet printed artificial neuron (AN) that performs the computing of a recurrent neural network when multiple units are coupled. An activation function (inverse hyperbolic sine) is fit to the electrical curve of the current-controlled graphene element with
$R$
-squared of 0.997. The standard neuron is replaced with the modeled one in the echo state network (ESN) for training on the MNIST benchmark dataset for handwritten digit classification. Testing performance of the simulated inkjet printed neuron reached 88.1% classification and was marginally better than the sigmoid and hyperbolic tangent functions. This work demonstrates a minimal-fabrication alternative computing element functioning as a simulated AN in the ESN. Benefits include low-cost fabrication, high power-efficiency, physically flexible edge computing, and development into many applications within telehealth, infrastructure, military, sports, etc. This work supports efforts toward a printable, cost-effective, flexible, and scalable physical machine learning system.