An electronic nose using a single graphene FET and machine learning for water, methanol, and ethanol

The study addresses the long-standing problem of poor selectivity in miniaturized gas sensors, particularly those based on metal oxide semiconductors (MOS), which often require high operating temperatures and suffer from cross-sensitivity. Demand for compact, low-power gas sensors is growing for applications such as air quality monitoring, occupational safety, consumer electronics, and medical diagnostics, where volatile organic compounds (VOCs) like methanol and ethanol are of interest. Electronic noses (e-noses) emulate biological olfaction using sensor arrays and pattern recognition to achieve selectivity, but typically rely on multiple functional materials and complex fabrication. The authors propose a new scheme that combines a single graphene field-effect transistor (GFET) with machine learning to achieve selective gas sensing at room temperature without multiple functional materials. The GFET’s V-shaped conductivity profile versus gate voltage is decoupled into four gas-informative physical properties, forming a 4D feature vector for pattern classification. The research aims to demonstrate that these vectors contain gas-specific information enabling selective identification of water, methanol, and ethanol, including in the presence of humidity.

Prior work on miniaturized gas sensors includes MOS sensors, which are sensitive and low-cost but operate at elevated temperatures (>200 °C) and have poor selectivity due to surface oxygen chemistry. Optical gas sensors offer high selectivity but are limited by size, configuration complexity, and cost. Electronic noses (e-noses), inspired by biological olfaction, have used arrays of cross-sensitive sensors and pattern recognition to improve selectivity since the late 1980s and 1990s, but broader commercial success has been limited. GFETs have been explored as room-temperature gas sensors with ultralow power consumption and distinct V-shaped conductivity profiles but also suffer from selectivity challenges. Previous studies examined individual physical properties (e.g., mobility, carrier concentration) for gas sensing, but not within a combined high-dimensional, machine-learning-based framework as proposed here.

Devices and gases: Two GFETs were used: a pristine GFET and an ALD RuO2-functionalized GFET (ALD-RuO2-GFET). Target analytes were water (H2O), methanol (MeOH), and ethanol (EtOH). Measurements were performed at room temperature. Measurement setups: Three setups (A, B, C) were configured. Setups A and B assessed repeatability and classification for individual gases; setup C probed binary mixtures (humidity with methanol). Common measurement apparatus and parameters were kept constant across experiments; variables were device type and gas type. Setup A results with a pristine GFET are emphasized in the main text, with additional results in Supplementary Information. Signal acquisition and feature construction: Time-resolved conductivity profiles (σ vs gate voltage VG) were recorded during controlled exposure cycles (e.g., ascending 10–90% and descending 80–10% gas concentrations). From each profile, four physical properties were extracted based on GFET transport models: - q1: electron mobility (μe) derived from slope of electron branch - q2: carrier concentration (n) inferred from neutrality point (VNP) shift - q3: hole mobility (μh) from slope of hole branch - q4: ratio of residual carrier concentration to charged impurity concentration (nr/Nimp) related to minimum conductivity at the neutrality point These properties are influenced by gas-induced charge transfer and Coulomb interactions at the graphene surface. The 4D vectors were normalized, and a sensitivity vector q(t) = 100 × (q(t) − q0)/q0 (%) was defined using the state prior to the first exposure cycle as reference. For visualization, a 3D vector set excluding carrier concentration (q2) was also used. Protocols and analyses: Gas-sensing patterns were evaluated for ascending-only cycles and for combined ascending/descending cycles (bounded by triangulated regions) in 3D feature space, with projections to representative 2D planes. Local repeatability (within a dataset) and global repeatability (across datasets/devices) were assessed. Pattern recognition and machine-learning analyses were used to classify the 4D/3D vectors by gas type. The inclusion of two device types enabled expansion of the feature space from 4D to 8D to test effects on classification accuracy.

- A single GFET, combined with feature extraction of four transport properties and machine learning, enabled selective classification of water, methanol, and ethanol vapors when tested individually, achieving high accuracy (qualitative claim based on large datasets). - Conductivity response was strongest for water and methanol; ethanol yielded smaller responses but still exhibited distinguishable patterns in the feature space. - 3D and 4D gas-sensing patterns showed consistent trends with good local repeatability; results from different setups and the ALD-RuO2-GFET indicated strong global repeatability. - Distinct bounded regions in feature space separated gas types, supporting qualitative classification without multiple functional materials. - Gas concentration dependencies were often near-linear, though some nonlinearities were observed (notably in ethanol’s carrier concentration changes), potentially due to interactions with pre-existing charged impurities. Despite nonlinearity, patterns remained distinguishable. - Extending the feature space from 4D (single device) to 8D (two devices: pristine and ALD-RuO2) further improved classification capability. - In binary mixtures (variable relative humidity with methanol), methanol patterns were qualitatively distinguishable from water vapor patterns, suggesting feasibility of identifying target gases under ambient humidity backgrounds.

The findings demonstrate that decoupling a GFET’s conductivity profile into four physically meaningful properties provides gas-specific signatures that, when analyzed via machine learning, overcome the traditional selectivity limitations of single-sensor systems. This approach leverages the intrinsic interactions between adsorbed molecules and graphene (charge transfer, Coulomb scattering, and impurity/residual carrier modulation) to encode patterns unique to each analyte. The robustness of patterns across devices and setups indicates strong repeatability and supports practical viability. The ability to qualitatively distinguish methanol signals in the presence of varying humidity is significant for real-world ambient sensing, where water vapor typically interferes with measurements. Adding a second, differently functionalized GFET expands the feature space and enhances separability, pointing to a scalable path for improved performance without resorting to large arrays of chemically distinct materials. While some property–concentration relationships showed nonlinearity, the overall pattern-based framework remains effective for classification and may be complemented by dedicated concentration estimation strategies.

This work introduces a miniaturized e-nose paradigm using a single GFET and machine learning to achieve selective gas sensing at room temperature without multiple functional materials. By converting GFET conductivity profiles into a 4D feature space (μe, n, μh, nr/Nimp), the system reliably distinguishes water, methanol, and ethanol, and qualitatively separates methanol responses from humidity in binary mixtures. The method exhibits good local and global repeatability and can be enhanced by incorporating additional devices (e.g., ALD-RuO2 functionalization) to increase feature dimensionality. Future work could focus on quantitative analysis of complex mixtures, robust concentration estimation (potentially via parallel GFETs optimized for quantification), algorithmic optimization for real-time classification, long-term stability and drift compensation, and expansion to broader analyte panels relevant to environmental monitoring and health diagnostics.

- The study focuses on three analytes (H2O, MeOH, EtOH); generalization to a wider set of gases and complex mixtures requires further validation. - Binary mixture analysis (humidity with methanol) demonstrates qualitative separation; comprehensive quantitative mixture deconvolution was not addressed. - Some gas–property relationships exhibited nonlinearity (e.g., ethanol’s carrier concentration), which may complicate straightforward concentration estimation. - Detailed machine-learning model specifications, training protocols, and quantitative classification metrics are not provided in the excerpt. - Device-to-device variability is only partially explored (two device types); broader manufacturing variability and long-term stability were not fully characterized.