Gas sensor arrays generate time-series responses to volatile organic compounds (VOCs), but traditional ML models ignore chemically relevant information of analytes and sensor functional group sensitivities. We propose a chemically-informed Transformer framework, combining molecular fingerprints (via RDKit), sensor functional group embeddings, and tubelet-based spatio-temporal patching in a ViViT-inspired architecture. The representation space encodes the structure and chemical nature of gases and sensor surfaces for VOC classification. Tested on methanol, propanol, butanone, formaldehyde, and their mixtures, the model captures signal patterns effectively and achieves 96.95% accuracy. This approach is interpretable and generalizable to other sensor datasets.