Scalable graphene sensor array for real-time toxins monitoring in flowing water

Inadequate drinking water management exposes millions to harmful contaminants, leading to diseases. The UN's Sustainable Development Goals aim for universal access to safe drinking water by 2030. Electronic sensors, with their rapid response, high sensitivity, low cost, and ease of integration with existing infrastructure, offer advantages over traditional methods like mass spectrometry for continuous online water quality monitoring. Two-dimensional (2D) layered nanomaterial-based field-effect transistors (FETs) show promise for detecting various pollutants, but scaling up their fabrication while maintaining quality control remains a significant hurdle. Device-to-device variations in response, calibration, and reliability hinder their widespread application. Existing approaches focus on controlling channel materials, but lack a holistic method to address device variations during large-scale manufacturing. This study presents a novel solution to address these challenges and enable the transition of 2D FET sensors from the research laboratory to real-world applications.

Numerous studies have explored the use of 2D nanomaterials, like graphene and MoS2, for chemical and biological sensing. Graphene-based FET sensors have demonstrated success in detecting heavy metal ions, gases, biomolecules, and bacteria. However, a critical challenge lies in scaling up the manufacturing process while ensuring consistent sensor performance across devices. While techniques like chemical vapor deposition, printing, spin-coating, and self-assembly have been explored for creating 2D nanomaterial-based sensors, they often lack effective quality control mechanisms to address device-to-device variability. This variability stems from variations in intrinsic and extrinsic properties of individual nanoflakes used in the FET architecture. The parallel connection of multiple flakes is explored to mitigate this issue by creating an ensemble effect but lacks a comprehensive method to effectively identify and address the issue of faulty devices before testing. This study aims to bridge this gap by introducing a comprehensive approach that addresses both large-scale fabrication and device quality control.

This study utilizes a bottom-up approach for the scalable nanofabrication of a graphene-based field-effect transistor (GFET) sensor array. Wafer-scale devices were fabricated through the wet transfer of single-layer graphene oxide (GO) dispersion and the patterning of interdigitated electrodes for parallel connection. A 80-nm SiO₂ layer was used as a protective layer for the electrodes, while a 3-nm Al₂O₃ layer served as the top-gate dielectric. Au nanoparticles were sputtered onto the Al₂O₃ surface as anchoring sites for probes. Three sensors were created using L-cysteine, thioglycolic acid (TGA), and an anti-*E. coli* antibody as specific probes for lead, mercury, and *E. coli*, respectively. The fabrication process resulted in ~60% of devices exhibiting narrow electronic property distributions. Faulty devices were identified by correlating bidirectional response behaviors with impedance measurements (Z`/Z`` > 1000 at low frequencies), attributed to defects in the Al₂O₃ layer. Drain current noise power spectral density (PSD) analysis further validated the absence of significant defects in near-ideal sensors. The responses of the GFET sensor array for simultaneous detection were calibrated using machine learning (ML) modeling, achieving high-precision classification and quantification at the ppb (cfu/mL) level. Impedance spectroscopy was used to characterize the quality of the 3nm Al2O3 dielectric layer. A model was developed using a superposition of two exponential functions representing the gating-induced current from analyte adsorption and defect-induced charge trapping. Noise spectral analysis examined the drain current noise power spectral density (PSD) to further evaluate the quality of the top gate dielectric layer and its interface with the sensor channel. The sensor array was integrated into a 3D-printed closed-loop chamber with a piezoelectric pump for continuous water flow testing, with sensors exposed to varying concentrations of heavy metals and *E. coli*. Machine learning, employing principal component analysis (PCA) and a two-layer feedforward artificial neural network (ANN), was used to classify and quantify the target analytes in mixtures.

The study successfully demonstrated a scalable method for creating GFET sensor arrays. A significant percentage (~60%) of the fabricated devices exhibited narrow distributions in resistance and on/off current ratios, indicating the effectiveness of the wet transfer and fabrication techniques in achieving uniformity. Impedance measurements and noise analysis proved effective in identifying faulty devices based on their non-ideal responses, with a Z'/Z'' ratio below 1000 at low frequencies serving as a practical criterion. The sensor array successfully detected lead and mercury ions and *E. coli* bacteria simultaneously in flowing tap water, achieving detection limits in the low ppb (cfu/mL) range. The sensor array's responses were successfully calibrated using machine learning, specifically principal component analysis (PCA) and a two-layer feedforward artificial neural network (ANN), leading to accurate classification and quantification of the target analytes even in mixtures. The ML model effectively compensated for cross-sensitivity observed in some sensors. Noise spectral analysis confirmed the chemical gating effect of analyte adsorption and provided further validation of the quality of the sensors' dielectric layer. The sensor array's performance was minimally affected by the pH and ionic strength variations in the tap water, highlighting its robustness and potential for real-world applications.

This study successfully addresses the critical challenges of scalability and device-to-device variability in 2D FET sensors for water quality monitoring. The combination of wet transfer techniques, impedance and noise analysis for quality control, and machine learning for data analysis provides a robust and reliable platform for real-time, continuous toxin detection in flowing water. The results demonstrate the feasibility of deploying such sensor arrays for early warning systems and improved water risk management. The high accuracy of detection, even in complex mixtures, showcases the power of the ML-assisted calibration. The minimal cross-sensitivity between sensors and robustness to pH and ionic strength changes highlight the practicality of this approach for real-world deployment. The findings have significant implications for advancing water quality monitoring technologies.

This research presents a significant advancement in the development of scalable and reliable 2D FET sensor arrays for real-time water quality monitoring. The combination of wet transfer fabrication, non-destructive impedance and noise-based quality control, and machine learning-based data analysis provides a robust solution to address the challenges of device variability and complexity in real-world water samples. Future research could focus on expanding the range of detectable toxins, improving sensor recovery times, and developing more sophisticated machine learning models to account for environmental factors such as varying pH and ionic strength.

The study focused on a specific set of toxins (lead, mercury, and *E. coli*). The generalizability of the findings to other contaminants needs further investigation. While machine learning effectively addressed cross-sensitivity, it relies on training data, and the accuracy might be affected by variations in water composition beyond the scope of this study. The sensor recovery time could be further optimized for improved continuous monitoring performance. The study employed tap water from a specific source; further testing with diverse water sources is necessary to validate the sensor's performance under varying conditions.