Miniaturized gas sensors are in high demand across various sectors due to their small size, low power consumption, and low cost. Metal oxide semiconductor (MOS) sensors are widely used but suffer from poor selectivity and require high operating temperatures. Optical sensors offer high selectivity but are typically large and expensive. Electronic noses (e-noses) aim to address the selectivity issue of MOS sensors by using arrays of sensors and pattern recognition algorithms. However, traditional e-noses require complex top-down functionalization processes. This research proposes a novel approach using a single GFET and machine learning to mimic the functionality of an e-nose. The GFET's unique V-shaped conductivity profiles, modulated by the interaction of gas molecules with the graphene surface, are used to extract four key physical properties. These properties are then used as a 4D feature vector for machine learning classification. The researchers chose water, methanol, and ethanol as target gases due to their importance as VOCs and the challenges posed by humidity in room-temperature GFET-based sensors. The study aims to demonstrate the feasibility of selective gas sensing using a single GFET without the need for multiple functional materials, paving the way for miniaturized, low-cost, and low-power e-noses.

The introduction thoroughly reviews existing gas sensor technologies, highlighting the limitations of metal oxide semiconductor (MOS) sensors (poor selectivity, high operating temperature) and optical sensors (size, cost, complexity). The authors discuss the concept of electronic noses (e-noses) as a solution to selectivity issues, outlining their components (sensor array, output vectors, pattern recognition algorithms) and the historical context of their development and limited commercial success. The use of graphene field-effect transistors (GFETs) as gas sensors is also reviewed, acknowledging their low power consumption and room-temperature operation but highlighting their poor selectivity. This review sets the stage for the proposed novel approach that combines the GFET's unique properties with the e-nose concept.

The researchers prepared two types of GFETs: a pristine GFET and an atomic layer deposition (ALD) RuO2-functionalized GFET. Three experimental setups (A, B, and C) were used to test the sensors with water, methanol, and ethanol vapors. Setup A, focused on in the main text, used a pristine GFET and involved exposing the sensor to individual gases while recording the conductivity profiles versus gate voltage over time. These profiles were converted into 4D vectors representing the four extracted physical properties. Setups B and C explored repeatability and binary mixtures, respectively. The 4D vectors were analyzed using machine learning algorithms to classify the gases. The sensitivity of the sensor to each gas was defined using a relative change in the 4D vector. Both ascending and descending concentration cycles were analyzed, creating 3D gas-sensing patterns that were visualized in 2D projections. The gas concentration dependence on each physical property was also analyzed. The paper includes equations for calculating the four physical properties (electron mobility, hole mobility, carrier concentration, and the ratio of residual carrier concentration to charged impurity concentration) from the measured conductivity profiles. Detailed information on the GFET fabrication process is provided in the Methods section of the paper.

The study demonstrated that the conductivity profiles of a single GFET exposed to water, methanol, and ethanol vapors could be converted into distinct 4D vectors representing four key physical properties. These vectors were successfully classified by machine learning algorithms, showing high accuracy in distinguishing between the three gases when tested individually. Furthermore, the system qualitatively distinguished methanol from water vapor in a binary mixture, indicating potential for operation in realistic scenarios with background humidity. The 3D visualizations of the gas-sensing patterns showed good local and global repeatability across different experimental setups and GFET types. Although some nonlinear relationships were observed in the gas concentration dependence of the physical properties, the qualitative distinction of gas-sensing patterns remained sufficient for successful classification. The results from setups B and C (presented in supplementary materials) further supported the robustness and generalizability of the proposed methodology.

The successful classification of water, methanol, and ethanol vapors using a single GFET and machine learning addresses the long-standing issue of poor gas selectivity in miniaturized gas sensors. The proposed approach overcomes the limitations of traditional e-noses by eliminating the need for multiple functional materials. The ability to distinguish between methanol and water in a binary mixture highlights the potential of the method for practical applications where background humidity is a factor. The findings suggest a new paradigm for gas sensing, leveraging the unique properties of GFETs and advanced pattern recognition techniques. The observed nonlinearities in gas-concentration dependence suggest that using multiple GFETs for parallel characterization could potentially improve gas concentration measurement accuracy, complementing the already effective selectivity achieved.

This study successfully demonstrates a novel gas-sensing scheme that uses a single GFET and machine learning to achieve high selectivity for water, methanol, and ethanol. The approach avoids the need for multiple functional materials, leading to miniaturization, low cost, and low power consumption. Future work could focus on exploring a wider range of gases and investigating the use of multiple GFETs for improved concentration measurement. The findings provide a promising pathway toward developing highly selective, miniaturized electronic noses for various applications.

While the study successfully demonstrated the feasibility of the proposed approach for water, methanol, and ethanol, further investigation is needed to evaluate its performance with other gases and in more complex gas mixtures. The nonlinear relationship between gas concentration and some physical properties might limit the quantitative accuracy of concentration measurements. The performance in real-world environments with various interfering gases requires further study. The reliance on machine learning may necessitate retraining the algorithm for different gas mixtures or environmental conditions.