Benchtop volatilomics and advanced convolutional neural network workflows for accurate and explainable food...

The analysis of volatile organic compounds (VOCs) by gas chromatography-ion mobility spectrometry (GC-IMS) offers high sensitivity, rapid analysis, and minimal sample preparation, but its second-order data complicate peak alignment, feature extraction, and classification. This study presents a deep-learning (DL) workflow using convolutional neural networks (CNNs) that treats GC-IMS chromatograms as images. Two CNN architectures simplified and complex were evaluated on olive oil and citrus datasets; the simplified model achieved 96.4 % and 98.3 % classification accuracy, respectively. To address limited sample sizes, a GC-IMS-specific data augmentation method that simulates plausible instrumental drift without chemically distorting signals was developed. For interpretability, saliency maps reveal discriminative features, and principal component analysis (PCA) synthesizes these into representative loading images per class, exposing stable collective VOCs signatures rather than single-sample explanations. By combining chemometric interpretability with DL prediction power, this framework provides a robust, explainable approach for GC-IMS data analysis for practical laboratory deployment.