Development of surface-enhanced Raman spectroscopy coupled with nanosubstrates and machine learning to improve food safety

The increasing use of pesticides in modern agriculture has raised serious concerns regarding food safety and public health due to the presence of hazardous residues in fresh produce. Traditional detection methods, such as chromatography, though accurate, are time-consuming, expensive, and require complex sample preparation. This dissertation introduces a novel and efficient approach for the rapid, non-destructive, and multiplex detection of pesticide residues by integrating surface-enhanced Raman spectroscopy (SERS) with advanced transformer-based machine learning models. This dissertation explores the development of novel SERS substrates through electrospinning. Electrospun polyacrylonitrile (PAN) nanofibers were functionalized and coated with Au@Ag core–shell nanoparticles to create high-surface-area scaffolds with strong plasmonic enhancement. These electrospun substrates demonstrated excellent uniformity and reproducibility, enabling the sensitive detection of the fungicide thiabendazole in soy-based foods with detection limits down to parts-per-billion levels (LOD = 23.1 ppb for soy milk; 79.4 ppb for soy sauce). The optimized electrospinning parameters produced nanofibers with a mean diameter of ~508 nm, offering a stable and tunable platform for SERS applications. This complementary approach highlights how nanofiber-based substrate engineering can significantly enhance the sensitivity and applicability of SERS sensors, further advancing the design of versatile detection platforms for monitoring food contaminants. Two innovative models, SERSFormer and SERSFormer-2.0, were developed and evaluated. Gold-silver core-shell nanoparticles (Au@Ag NPs) were synthesized to serve as SERS substrates, offering enhanced Raman signal amplification and reproducibility. The SERSFormer model was designed for single-pesticide detection in spinach samples and achieved 98.4% classification accuracy with a mean absolute error (MAE) of 0.966 in quantification. Building upon this foundation, SERSFormer-2.0 addressed the complex challenge of detecting multiple co-existing pesticide residues in real-world produce such as strawberries and spinach. This model employed a multitask attention-based transformer architecture capable of performing both multilabel classification and multiregression simultaneously. It achieved near-perfect performance across all metrics (accuracy = 0.999; F1 score = 0.992; precision = 0.990; recall = 0.996) and demonstrated strong regression capabilities with an MSE of 0.136 and R² = 0.804. These models utilize shared self-attention mechanisms and task-specific embeddings to capture complex spectral features, even in the presence of overlapping Raman signals. Preprocessing techniques, including denoising, baseline correction, normalization, and genetic algorithm-based feature selection, further enhanced model robustness. The comprehensive framework developed in this work not only improves detection sensitivity and accuracy but also allows real-time assessment of contamination in complex food matrices. This dissertation demonstrates that the synergistic combination of SERS and deep learning, particularly transformer architectures, offers a transformative solution for food safety monitoring. The findings have strong implications for agricultural management, public health, and the development of portable, AI-powered sensors for rapid on-site pesticide screening.

Name thesis:degree_name

Ph. D.

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Food and Hospitality Systems

Grantor dc:publisher

University of Missouri--Columbia

Year dc:date.issued

2025

Author dc:creator

• Hajikhani, Mehdi

Advisor dc:contributor.advisor

• Lin, Mengshi

Language dc:language.iso

eng, English

DOI

https://doi.org/10.32469/10355/111000

OAI identifier oai:identifier

oai:mospace.umsystem.edu:10355/111000