Physical Virtualization of a GFET for a Versatile, High‐Throughput, and Highly Discriminating Detection of Target Gas Molecules at Room Temperature

Abstract

An e‐nose is built on a single graphene field effect transistor (GFET), based on a graphene/Si3N4/p‐Si stack of layers. Multichannel data acquisition, enabling to mimic the architecture of a sensor array, is achieved by steering the gate potential, thus yielding a virtual array of 2D chemiresistors on a single sensing layer. This setting allows for the detection of volatile compounds with a remarkable discrimination capability, boosted by intensive machine learning analysis and accuracy maximization through the choice of the number of virtual sensors. Sensing of gas phase NH3 is tested, along with a set of possible interferents, and discrimination of NH3+NO2 mixtures is successfully probed. High throughput in terms of sensitivity is achieved by tracking the shift of the minimum of the GFET transfer curve versus NH3 concentration. With this readout scheme, a 20‐fold sensitivity increase over a 5–50 ppm range is registered to the same layer used as a chemiresistor. High discrimination capability is probed by leveraging machine learning algorithms, from principal component analysis (PCA) to Uniform Manifold Approximation and Projection (U‐MAP) and, finally, to a Deep Neural Networks (DNN) where input neurons are the virtual sensors created by the gate voltage driving. For the tested case, the DNN maximum accuracy is achieved with 21 virtual sensors.