Access to clean water is a critical global issue, with 2 billion people lacking safely managed water services. Organic pollutants pose a significant threat due to their persistence and toxicity. Traditional water purification methods have limitations in efficiency, speed, and potential for secondary contamination. Zinc oxide (ZnO), known for its photo- and piezo-catalytic properties, offers a sustainable alternative. However, enhancing ZnO's efficacy and preventing nanoparticle pollution requires careful material design. Monitoring water quality is equally crucial, and while techniques like SERS offer high sensitivity and speed, analyzing complex Raman spectra from mixtures of pollutants remains challenging. Machine learning (ML) algorithms, particularly deep learning (DL), can address this by extracting complex features from spectral data. This study aims to develop a dual-functional material system using Ag nanoparticles decorated ZnO nanorods coated silica nanofibers (AgNP-ZnONR-SNF) for efficient pollutant degradation and sensitive, quantitative detection using SERS combined with a novel DL algorithm for accurate qualitative and quantitative analysis, including out-of-distribution sample detection.

The introduction extensively reviews existing water purification techniques (chemical precipitation, filtration, adsorption, etc.) and their shortcomings. It highlights the advantages of ZnO's photo- and piezo-catalytic properties for water treatment, while also noting the challenges in optimizing ZnO nanostructures for improved degradation efficiency and reusability. The review also covers existing water quality monitoring techniques (photoluminescence spectroscopy, mass spectrometry, HPLC, SERS) and the challenges in detecting complex mixtures of organic pollutants. The application of various machine learning algorithms (PLS, support vector machines, convolutional neural networks, deep learning) to analyze SERS data is discussed, along with the limitations of current ML-assisted Raman detection methods, particularly in handling mixtures and out-of-distribution samples.

The study involves three main stages: material synthesis, pollutant degradation testing, and SERS detection with ML analysis. **Material Synthesis:** Silica nanofibers (SNFs) were fabricated via electrospinning. ZnO nanorods (ZnONRs) were grown on the SNFs using a seeding-growth hydrothermal method. Finally, silver nanoparticles (AgNPs) were decorated onto the ZnONR-SNFs via UV irradiation. **Pollutant Degradation:** The photo- and piezo-catalytic activities of the AgNP-ZnONR-SNF film were evaluated using methylene blue (MB), trypan blue (TB), and methyl orange (MO) as model organic pollutants. Degradation efficiency was measured via UV-Vis spectrophotometry under UV irradiation and shaking conditions. Reusability was tested over six cycles. **SERS Detection and ML Analysis:** SERS measurements were performed using a 532 nm laser. A deep neural network was developed for qualitative and quantitative analysis of the Raman spectra. The network incorporates a Laplacian operation to enhance peak information. The network outputs both a classification (identifying components) and a regression (quantifying concentrations) model. A k-Nearest Neighbors (KNN) model was added to detect out-of-distribution samples (unseen pollutants). The model was trained and tested using mixtures of MB, TB, and ciprofloxacin (Cip) at varying concentrations. The performance was assessed using accuracy, specificity, and absolute error metrics.

The AgNP-ZnONR-SNF thin film exhibited excellent photo- and piezo-catalytic degradation of MB, TB, and MO, achieving >98% efficiency under UV irradiation. The AgNP decoration significantly enhanced SERS signals, resulting in a 1056 enhancement factor and a detection limit of 1 pg mL⁻¹. The intensity of characteristic peaks showed a strong linear correlation with the concentration of MB and Cip. The ML algorithm enabled accurate (92.3%) qualitative and (90.8%) quantitative detection of mixtures of MB, TB, and Cip. The integration of a KNN model improved the detection of unseen classes with 80% sensitivity and 89.3% specificity. The Laplacian operation effectively enhanced the informative content of Raman spectra, improving overall performance. The uniform Raman signal intensity across the AgNP-ZnONR-SNF chip ensured consistent sensing performance.

The results demonstrate the successful development of a dual-functional nanophotonic sensor for both water purification and pollutant detection. The high efficiency of the photo- and piezo-catalytic degradation combined with the ultrasensitive SERS detection and accurate ML-based analysis showcases the potential of this technology for practical applications in environmental remediation. The ability to detect and quantify multiple pollutants simultaneously from complex mixtures, without extensive sample preparation, is a significant advancement. The inclusion of the KNN model enhances the robustness of the system by allowing the detection of unexpected pollutants, increasing its reliability for real-world scenarios. The cost-effectiveness and reusability of the sensor further enhance its practicality.

This study successfully demonstrated a novel dual-functional nanophotonic sensor for simultaneous water purification and highly sensitive organic pollutant detection. The integration of advanced nanomaterials and machine learning algorithms provides an efficient, accurate, and cost-effective solution. Future research could focus on expanding the library of detectable pollutants, optimizing the ML model for even higher accuracy, and exploring the application of this technology in various real-world water purification and monitoring settings.

The study primarily focused on a limited set of model organic pollutants. The generalizability of the ML model to a broader range of pollutants needs further investigation. While the sensor demonstrated good reusability, long-term stability and effects of fouling require further study. The current methodology for out-of-distribution detection relies on a KNN model trained with a specific set of unseen compounds, so further investigation might be required to enhance the robustness to more diverse unseen substances.